Climate Change AI

The Virtual Climate Change AI Summer School: 2024 Recap

2025-02-07T00:00:00+00:00

This past summer, from June to August, thousands of virtual CCAI Summer School attendees from over 150 countries expanded their knowledge across a broad range of topics at the intersection of AI and climate change, engaging with over twenty hours of expert-led content, working on curated tutorials and interacting on the CCAI Community Platform. These materials are now publicly available on YouTube and other channels - find all the links on our website!

In August, we also hosted an in-person CCAI Summer School in Mila - Quebec AI Institute in Montreal, Canada. You can read about the highlights in this blog post.

The virtual CCAI Summer School aimed to democratize education at the intersection of climate change and machine learning by making high quality learning experiences accessible across the globe. To that end, the program could be completed from anywhere with internet access and allowed for both live and asynchronous participation. The program was targeted at both beginners and experienced professionals with backgrounds in either climate change related fields or machine learning, but not necessarily in both areas. To help those who had little to no machine learning experience, we developed a set of publicly accessible introductory learning materials that can be found here. Also, the lectures and tutorials were presented at varying levels of technical depth, so that every participant would be able to take something away from each session.

“There was such a fantastic breadth and depth to the topics covered, I honestly didn’t expect to deep dive into such discrete topics such as agriculture, transportation, buildings. etc.. I feel like being able to do that gave me such a better and wider look at AI then what we are currently seeing in the media/online. I liked that the topics weren’t surface level but very deep with specific examples and case studies”

– Anonymous Virtual Summer School Participant

Kai Jeggle virtually presents a slide on AI and Climate

The curriculum provided an introduction to the space by showcasing many different application areas and addressing overarching topics such as policy and impact assessment. The first day of the Summer School included lectures that covered introductory topics related to climate change, machine learning, and the intersection of these fields. From there, we dove into a variety of topics, including agriculture, climate science, ethics of AI, transportation, health, and more. Some lectures built upon others to provide more specific deep-dives, and some were accompanied by tutorials that provided a hands-on learning experience. For example, Dr. Kai Jeggle provided an overview of AI for climate science, which was followed by a lecture from Dr. Peetak Mitra specifically focused on the application of AI for weather. After the lectures, participants had the opportunity to apply the concepts from these lectures to a tutorial on forecasting El Niño with machine learning authored by Ankur Mahesh and Mel Hanna.

Marcus Voss presents a virtual live runthrough of the Building Load Forecasting with ML tutorial

Even though this was a virtual program, there was plenty of opportunity for interaction. Participants engaged directly with experts in the live lecture Q&A sessions, and with each other through the CCAI Community Platform and social events.

“The program has bolstered my confidence in pursuing a career that intersects AI and climate change. The diverse range of topics and expert perspectives has reinforced my commitment to working in this impactful area.”

– Wasim Saeed Hasan Hezam, Virtual Summer School Participant

In addition to the lectures, attendees gained practical experience through the tutorials that accompanied most of the lectures. These hand-on materials address problems such as reducing the climate impact of training ML Models, modeling for bioacoustic monitoring, and using language models to analyse climate policy. The tutorials remain public after the end of the summer school, and provide an excellent opportunity to explore datasets, models and case-studies, and get hands-on experience with implementing ML models in this space. All the tutorials are available on the CCAI website.

All content from virtual Summer School is now accessible on our YouTube channel and the Summer School website.

The In-Person Climate Change AI Summer School: 2024 Recap

2024-12-04T00:00:00+00:00

The climate change and AI space is growing rapidly, requiring democratization of expertise on these topics as well as the development of leaders who are able to shape the space responsibly. In August, we brought together 34 future leaders to the Mila - Quebec AI Institute in Montreal to spend a week together expanding their skills in the field. During this time, participants engaged with training sessions on project development skills, touching on topics such as pathways to impact and responsible AI. Their learnings were immediately put to use: under the guidance of expert mentors, participants worked together in teams to develop projects using AI to solve various climate related challenges. The summer school also focused heavily on building lasting relationships among participants through social and networking events. Keep reading for the 2024 Summer School highlights, a closer look into each team’s project, and how to get involved with the next edition of the program.

A Diverse Cohort

CCAI’s In-Person Summer School is a project-based program geared towards individuals with background in AI and/or a climate change-related field, across many career stages. We received over 850 applications for this year’s program and were excited to be able to bring 34 of those applicants to Montreal. This year’s cohort was curated to span across careers and areas of knowledge: participants ranged from activists, to researchers, to government officials, with expertise in topics such as hydrology, policy-making and data science, and came from over twenty five countries.

“Throughout the summer school, I had the opportunity to connect with a passionate and talented community of individuals dedicated to leveraging AI for climate action. The relationships forged and insights gained will undoubtedly lead to exciting collaborations and collective impact in the future. Overall, the CCAI Summer School has been a catalyst for my growth as a climate advocate and has equipped me with the knowledge and network to make meaningful contributions at the intersection of climate and AI.”

– Nabila Putri Salsabila, In-Person Summer School Participant

The Summer School was a fantastic opportunity for participants to gain hands-on experience working with an interdisciplinary team on an AI-for-climate project, supported by workshops and discussion sessions that enabled participants to produce more impactful projects, become thought leaders and connect as a cohort. Our instructors delivered training sessions on skills that are important for work at the intersection of climate change and machine learning: pathway to impact, responsible AI, working effectively in interdisciplinary teams, and effective communication. The participants took the learnings from these sessions and applied them to inform the responsible development of their projects and create compelling pitches. For example, the sessions led to one team researching the uses of AI to enhance satellite imagery to explore the potential ethical drawbacks of their proposed solution.

Jesse Dunietz delivers a session about effective communication to the Summer School attendees

The summer school was not only work – there was fun too. For instance, during a 5à7 at Mila, participants got to know each other better and networked with some members of the Mila ecosystem. Participants also got to visit Montreal’s TimeOut Market, and explored the city together.

Project Development

Participants work together in groups at Mila - Quebec AI Institute

Participants formed ten interdisciplinary teams based on project proposals that they developed individually. Even before they arrived in Montreal, teams scoped out key aspects of the project together, such as the potential impact on climate change of the project and the stakeholders who may be affected by its implementation. They also identified the datasets necessary for their projects and how they would prepare those datasets to use during their project. With all of this prep work done, each team was able to hit the ground running when they arrived in Montreal. Outside of training sessions and social events, participants spent the majority of the week working with their teams. Supported by expert mentors, each team put significant effort into developing their projects and preparing for the final deliverable: the pitch event.

Pitch Event

After a week of hard work, each team was prepared to present their project to the rest of the cohort and a panel of expert judges. We heard from one team about increasing the lifespan and efficiency of photovoltaic modules by forecasting energy production, detecting degradation patterns, and predicting module failures. Another team presented their own flood prediction models in Kenya, which would provide an early warning system and potentially help Kenya to build a national flood management system. One team even demonstrated their climate vulnerability index through a live interactive demo! At the bottom of this page you will find a description of all the projects the participants worked on. And the work does not end with the Summer School: multiple teams submitted and had their projects accepted to the Tackling Climate Change with Machine Learning workshop at NeurIPS and may continue to develop these projects further.

Summer School participants pitch their project

Goodbye, For Now

This year’s In-Person Summer School was truly an enriching experience for everyone involved. The participants’ drive for developing solutions for climate related problems using machine learning was inspiring, and their connections and work will extend far into the future. Before parting ways, many of us gathered for one final dinner, where we enjoyed great conversations and celebrated everyone’s terrific work. It was a great way to wrap up the program and say goodbye - at least, for now.

Selfie at the closing dinner of the Summer School

Interested in participating in the next iteration of the In-Person Summer School? Sign up here to be notified when our next application call is released.

Project Descriptions:

Lifespan and Efficiency of Solar Photovoltaic Modules
Giulia Lombardi, Andrés F. Pérez, Alvar Herrera, Christian Tran
The sustainability of future energy relies on the performance of photovoltaic (PV) modules, often hindered by degradation. This research seeks to improve the lifespan and efficiency of PV modules by using artificial intelligence (AI) and machine learning (ML) to develop predictive models based on historical performance data. These models will identify degradation sources affecting energy production, enabling optimized maintenance schedules, cost reductions, and minimized environmental impact. Specifically, we aim to use existing datasets to forecast energy production, detect degradation patterns, and predict module failures, supporting proactive maintenance and enhancing the reliability of solar power.

Downscaling For Improved Precipitation Forecasts In Data-Scarce Regions
Miriam Simm, Rajesh Kandel, Rogelio Barrios, Akram Zaytar
Climate change is expected to increase the frequency and intensity of extreme precipitation events, posing significant threats globally. Accurate and reliable precipitation forecasts are crucial for mitigating these impacts. While Numerical Weather Prediction (NWP) models provide global forecasts, their coarse resolution limits the accurate prediction of precipitation. Machine Learning (ML) based downscaling methods have emerged as a promising way forward,
though their application in data-scarce regions remains underexplored.

This study aims to develop an ML-based model in order to improve precipitation forecasts for the Karnali river basin in western Nepal, a data-scarce region vulnerable to extreme weather events. By training on high-resolution NWP simulations and fine-tuning with local observations, we aim to enhance forecast accuracy and reliability. The approach’s generalizability will be evaluated on other data-limited regions, with the goal of creating a flexible downscaling toolbox for broader application.

Climate Vulnerability Compass: Leveraging Machine Learning for Climate Vulnerability Assessment and Actionable Recommendations
Flávio Nakasato Cação, Isabel Drummond, Donna Marie Mlyneck
This paper proposes a framework that utilizes climate vulnerability indices, large language models (LLMs), and information retrieval (IR) systems to generate actionable recommendations for mitigating and adapting to potential environmental disasters. By integrating quantitative data from pre-existing indices with qualitative insights from policy documents, this project aims to equip public managers, journalists, and civil society with practical tools for understanding and addressing climate risks in specific regions. The ethical implications and pathways to impact of this approach are also discussed. Our prototype is available here: https://github.com/nakasato/climate-vulnerability-compass

Flood Prediction in Kenya
Alim Karimi, Hammed A Akande, Valerie Brosnan, Nicole Mong’are, Asbina Baral
We aim to build a flood prediction model which can be used by Kenyan authorities to deploy an early warning system. We aim to build two models. The first model will be predicting flooding events within a 900 square kilometer region. The second model will predict flooding events on a
much coarser 900 square meter region. Given that Kenya does not have a national flood management system, we hope that these technical advances can aid in building a national flood management system in Kenya and also in other countries which lack the infrastructure.

Advancing Climate Resilience in the Global South Through Enhanced Satellite Imagery
Zhining Gu, Muhammad Zawish, Eloise Lopez, Briah Davis
While High Resolution (HR) Remote Sensing (RS) imagery plays an important role in advancing climate resilience in various applications, such as damage estimation, crop yield estimation, and farm boundary delineation, its acquisition is prohibitively expensive. Fortunately, the cutting-edge Super Resolution (SR) techniques provide low-cost solutions to generate HR images from Low Resolution (LR) data. In the field of RS, due to the availability of ground-truth data (HR images) and sufficient financial support, SR models are frequently applied to the global north, but limited to developing countries across the global south. As a result, the improvement of climate resilience (i.e., disaster response, agriculture monitoring and management) in the global south is largely restricted.
To solve this issue, we leverage Open Data Program launched by MAXAR, one of the largest RS imagery producers, to curate HR and LR image pairs for the global south as benchmark datasets. The evaluation on various state-of-the-art SR models will be conducted. With the trained models, we further apply them to various downstream tasks. Finally, this project promises to improve the accessibility to HR satellite imagery for developing countries in the global south, assisting in various climate-related applications, such as disaster response, farm boundary delineation, crop type classification, and so on.

Understanding Crop Dynamics in Kenya
Hannah Wang, Nabila Putri Salsabila, Nirdesh Sharma
Kenya’s agriculture is vital, supporting much of the economy and employing a large part of the population. However, it faces serious challenges from climate change and food insecurity. This project aims to improve the health of Kenya’s soil by predicting soil organic carbon (SOC), a key indicator of soil fertility, using advanced data techniques like satellite imagery and machine learning. We start by identifying areas where crops are most likely grown using global land cover data. Then, we develop a machine learning model that not only maps crop types across the country but also includes a feature that measures the certainty of its predictions. This means the model can flag areas where it is less confident, helping us focus on getting the most accurate results. By combining these detailed crop maps with other data, such as soil and climate information, the project has already improved the accuracy of SOC predictions. Ultimately, this will help farmers adopt better farming practices, guide policy decisions, and contribute to global efforts against climate change, all while ensuring the benefits reach everyone, especially marginalized communities.

Monitoring the evolution of glacier surface area in the Peruvian Andes
Codrut-Andrei Diaconu, Gabriela Yaulli
Our project aims to develop a fully automatic pipeline to monitor the evolution of glacier surface area in the Peruvian Andes over a specific period of time. The goal is to provide annually updated area estimates for each glacier in the region, together with a tool that can be used by researchers who build glacier inventories to simplify their work. As a major part of our project, we plan to create a dashboard to visually represent changes in glaciers over time. We also aim to incorporate existing research findings that project how glaciers will change by the end of the century. This will provide a comprehensive view of both recent changes and potential future scenarios. The dashboard can be used as printed information for tourists and can also be made accessible on the web. This web version can provide continuously updated results and use animations to enhance the presentation.

From insight to action: Empowering residential users to lower their carbon emissions
Shivani Chotalia, Kyungmin Lee, Victor Hutse, Soazig Kaam
Residential buildings account for 17-20% of global greenhouse gas (GHG) emissions, posing a significant challenge in the effort to transition all buildings to net zero. While solutions such as heat pumps, building retrofits, and load shifting can significantly reduce emissions, homeowners are often unaware of their energy usage and the potential for cost reduction. Current methods for assessing individual building energy use typically rely on invasive approaches such as walk-throughs, surveys, and in-situ measurements, which can pose challenges related to data access and technology literacy, particularly for users. Supervised machine learning methods have the potential to provide actionable insights for individual energy-end users, encouraging them to reduce both GHG emissions and energy costs. We propose a novel framework that utilizes baseline, intervention, and optimization models to predict emissions and cost estimates for individual energy-end users. This paper presents a novel application of an optimization model for energy bills through machine learning methods: (1) classification of time series data for electricity and gas usage baselines, (2) prediction of GHG emission reductions, and (3) prediction of energy cost reductions.

Helping global grids work smarter, not harder

2024-08-29T00:00:00+00:00

Electric grids are astonishing achievements of human ingenuity. When you casually plug your phone into a wall outlet, the electric power that flows to charge it has reached you through a vast and sprawling network of wires—some strung between towers, others buried; some fine, others thick and coiled—that connect you to many distant and diverse generators where energy from wind, sun, and fuel is transformed to electric power. This vast network, or electric grid, is legitimately described as the world’s largest machine: all of the grid’s seemingly disjoint elements are in fact intimately connected, operating in precise and carefully controlled unison. When you plug in your phone, each distant generator feels the minute increase in aggregate system load. In turbine-based generators (such as conventional fuel powered ones), this increase causes the turbine’s rotation to slow fractionally; when the turbine’s slowing exceeds some operator-defined limit, a control loop running in the generation plant increases the fuel supply to meet this new demand, speeding the turbine back up. As millions of us around the country and world power our devices, the unified pulse of generator turbines allows the grid to maintain a perfect balance of supply and demand, providing each of us with the perfect amount of electricity, rarely too little, rarely too much, to go about our day.

Despite the physical marvel of the synchronous grid, many of us are so accustomed to electrical grid infrastructure that we take it for granted. When did you last notice the wires that hung from your home, connecting it to nearby utility poles? Or take note of transmission lines cutting across the country while on a road trip? The electrical grid forms the backbone of our society, and yet we rarely acknowledge it except when it isn’t working as expected. While it might be nice to have your eyes subconsciously filter out the messy electrical lines in a sunset view, our society “not noticing” the grid has considerable detrimental implications in the current historic moment, when we are demanding more from our grids than ever before.

Today, electrical grids around the world are facing an unprecedented dual challenge. Climate change mitigation requires that we decarbonize energy production, thus fundamentally changing the sources and sinks of electricity carried by our grids and pushing them to support larger, more complex electricity flows. Simultaneously, we must combat global energy poverty: as United Nations Sustainable Development Goal 7 enshrines, providing access to “affordable, reliable, sustainable and modern energy for all” is fundamental to alleviating poverty and human suffering. Combating global energy poverty requires hundreds of millions of additional grid connections, pushing grids to operate closer to their limits and to reach more remote populations.

A gap in energy usage. Electricity use per person, in kilowatt-hours (kWh), per country around the world (source: Our World in Data). LMICs, especially in regions like sub-Saharan Africa, contribute little to the world’s energy burden today. However, to raise these regions out of energy poverty necessitates demand growth that will place LMICs more on par with high income countries. Meeting this demand, while simultaneously realizing global decarbonization goals, is one of the most pressing challenges facing the world today.

These twin challenges—decarbonization and ending energy poverty—are daunting everywhere in the world, but will fall most heavily on grids in low- and middle-income countries (LMICs) where decades of limited infrastructure investment have produced weak and inadequate grids. These weak grids are already being asked to operate beyond their limits; poor power quality and reliability (PQR) originating from weak or undersized grids affects an estimated 3.5 billion people in LMICs. In these same LMICs, electricity demand is set to grow swiftly, climate impacts are projected to be severe, and electrical utilities often lack the capacity and tools available to their counterparts in high-income countries (HICs) to strengthen their grids in the face of these challenges.

Smart grids are useful…and expensive

One critical tool often lacking on the grids of LMICs is smart-grid technology: sensors and control equipment deployed on the grid that can automatically monitor the grid’s state and take action to maintain grid stability.

In HICs, smart grid technology has been a critical enabler of the energy transition: decarbonized electricity generation from solar or wind technology leads to more volatile and unpredictable energy flows, and sophisticated monitoring and control can ensure that this electricity reaches customers seamlessly, with no gaps, and at stable levels that do not damage grid equipment or consumer appliances.

However, conventional smart-grid technology is costly, slow to deploy, and can be difficult to use. Existing state-of-the-art sensing technologies can cost from $40,000 to $300,000 USD [1], [2] per deployed device; smart-meter rollouts in the EU cost €40B [3]. Additionally, transforming smart-grid data into actionable insights demands technical capacity that is absent for many utilities and regulators in LMICs. It is unclear whether traditional smart-grid technologies will be a viable solution in LMICs on the timescales required to realize the global energy transition and support climate change mitigation and resilience.

The GridWatch Dataset in Accra, Ghana

It is deeply unjust to demand that LMIC utilities accomplish the unprecedented—raise growing populations from energy poverty while transitioning to clean energy with minimal resources—without substantial global support in developing the technological, social, and structural innovations needed to do so. However, most technology development and research for smart grids today is done within HICs contexts. Through the grant award of CCAI’s 2023 Innovation Grants Program, we hope to lay a foundation for more smart grid development to be contextualized within the needs and constraints of LMICs.

We are building this foundation with three pillars:

GridWatch: An alternative smart-grid sensor. Through the CCAI Innovation Grant, we built a partnership between power and energy systems professor June Lukuyu at the University of Washington and nLine a small social enterprise that develops and builds low-cost smart grid technologies designed specifically for LMIC contexts. nLine’s GridWatch sensor plugs into a standard outlet and collects data about power reliability and quality. These data can be used to detect failing equipment, restore outages more quickly, diagnose systemic challenges in the grid, and understand the human experience of electricity. Sensors report data over the cellular network using a universal SIM, store it locally in times of poor network quality, and include a battery so they are ready to collect data even after long outages. This sensor is 1/100th of the operational cost of alternative smart grid sensors, and because it plugs directly into a wall outlet, it can be deployed and maintained directly by a non-electrical expert to provide rapid, robust insights.
Open-sourced data from 1,300 grid sensors in Accra Ghana. nLine deployed 1,300+ GridWatch sensors in Accra, Ghana over 4 years, in a project funded by the Millenium Challenge Corporation and the Government of Ghana. As part of the CCAI Innovation Grant, we are thrilled to release this dataset publicly via a custom-built dashboard that allows easy visualization, access, and download. The Accra Power dataset represents the first, large-scale longitudinal study of ground-truth power quality and reliability on the electrical grid in a LMIC. It provides a rich foundation upon which to develop and prototype smart grid solutions. Voltage data and grid frequency are collected every 2 minutes; these raw data are used to detect outages and compute grid key performance indicators, such as SAIDI, SAIFI, and hours undervoltage. Both raw data and KPI values are available as part of the open-sourced Accra Power dataset. We are excited to bring this dataset to the community of energy systems researchers and practitioners; please reach out to info@nline.io if you have any questions about the dataset.
Machine-learning-based smart-grid use cases. To make grid data valuable to capacity-constrained utilities, we are building out a set of canonical smart-grid use cases, which demonstrate how smart-grid data can be used to help utilities reduce outage duration and technical loss in their network. As part of the CCAI Innovation Grant, we are developing an ML-based topology reconstruction algorithm to support these smart-grid use cases.

A call to help global grids work smarter, not harder

Electric grids are the conduits through which the majority of clean, modern energy will be transmitted, and grids will play a central role in addressing the linked challenges of climate mitigation, adaptation, and resilience. To enable this critical public infrastructure to rise to the new and unique challenges of our time, we must invest in grids around the world, and especially in LMICs, which are being asked to abate and face the climate impacts they did little to cause without adequate assistance. Without novel approaches to making grids smarter, LMIC grids may remain too weak to meet climate challenges. This will compromise climate mitigation and adaptation strategies (many of which depend on good electricity supply), leaving vulnerable populations even more exposed to the impacts of climate change. Leveraging the tools and datasets developed in this project, we hope to see a new generation of smart-grid algorithms that are more accessible, lower cost, and designed for LMIC contexts.

If you are a researcher developing smart grid algorithms, an electricity provider, a donor funding electrification projects, or otherwise involved in this field, we invite you to share your suggestions and feedback on how the Accra Power dataset could support your work. We are also keen to hear your ideas on how we can enhance our datasets and algorithms to better enable smart-grid applications in LMIC contexts.

The GridWatch team. Our team meeting in Dakar, Senegal, to reflect on plans to make grid technology work more just and accessible around the world.

The Climate Change AI Innovation Grants Program 2023 is supported by Quadrature Climate Foundation, Google DeepMind, Global Methane Hub, and the Canada Hub of Future Earth.

AI for Public and Planetary Health

2024-08-09T00:00:00+00:00

Dr. Sara Khalid is an Associate Professor and Head of the Oxford Planetary Health Informatics Lab in the Centre for Statistics in Medicine at Oxford University. She is also a Senior Research Fellow in Biomedical Data Science and Health Informatics and a National Geographic Explorer in Remote Monitoring and Machine Learning.

She talks here about her work in climate and planetary health with CCAI’s Public Health Lead, Sarah Skenazy. You can hear more from Professor Khalid in the recording of her recent CCAI Summer School lecture on climate and health. Sign up for our newsletter to be alerted to the release of the session!

Interview with Dr. Sara Khalid

Skenazy: What is the relationship between climate change and health—how is climate change impacting global public health and what can we anticipate happening in the future?

Khalid: Climate and health are inextricably linked, both directly and indirectly. In fact climate change has been recognised as one of, if not the, greatest public health challenges of the 21st century.

Why is that? We know that the environments we are exposed to—natural, living, and social—lead to the majority of diseases, both chronic and communicable. The World Health Organisation (WHO), which has been monitoring this for a long time, estimated that one in four deaths around the world are linked to the environment. Climate change is multiplying this threat. Take the example of extreme weather – extreme heat and cold temperatures are associated with more than 5 million excess deaths every year. On top of the human cost, climate change has a significant economic cost, with WHO estimates suggesting over $4bn/year.

As the Earth’s vital signs continue to steer off course, climate-induced disasters, biodiversity loss, and extreme weather events are on the rise around the world. Heatwaves, for example, are becoming longer-lasting, more intense, and more frequent, and are amongst the deadliest extreme weather events. Within the last decade, extreme heat events have triggered public health red-alerts globally. Over 30 countries have declared climate emergencies. Moving forwards, there is an imminent and growing need to adapt our health solutions and care systems so the health of everyone, especially those most vulnerable, is protected.

Skenazy: Your work with the Planetary Health Informatics Lab sometimes frames the interdependent risks facing the earth and all beings on it in terms of a ‘planetary health crisis.’ Can you explain how planetary health relates to global public health, health equity, and related frameworks (like climate justice) across the public and population health field internationally?

Khalid: Planetary Health is simple in concept, it broadly refers to the study of the impact of human-induced disturbances in the planet’s ecosystem on human health itself. So the focus is still very much on human health and its intersectional determinants, although the causes are not limited to traditional physiological and genetic risk factors but rather the entire exposome, which encompasses all environmental exposures that an individual encounters throughout their life, and how these exposures impact both biology and health.

Climate justice is important because not everyone is equally affected by climate change. Within a given society, for instance, evidence indicates that the elderly, children, pregnant people, outdoor workers, and unhoused people are disproportionately at higher risk. The recent ruling by the European Court of Human Rights in favour of elderly women highlights the importance of recognising a link between human rights and climate change.

But there is still much we don’t fully understand. Specific knowledge gaps include how individual-level health outcomes of at-risk groups are affected, and how these impacts vary with factors like socio-demographics, genetics and/or ethnicity, and different climatic regions. These gaps can transfer into health adaptation practices and policy-making.

On another level, although economically rich countries have historically generated more greenhouse gas emissions, it is some of the most economically poor regions that have some of the highest climate vulnerability indices. The same under-resourced countries often have overburdened fragile healthcare systems, and so that leads to injustice in public health and health inequity more broadly. On the flip side, countries where people may be less acclimatised for instance to extreme heat or are historically less exposed to infectious diseases may be inequitably at risk because of a lack of both risk awareness and supportive infrastructure. So it is a multi-faceted complex global issue.

Skenazy: What is the role of artificial intelligence in addressing the planetary health crisis?

Khalid: In the field of healthcare, AI can be used as a force for good. One of its most successful applications has been in traditional medical imaging, by helping doctors to see what the naked eye cannot. In this way, it is helping treat cancers, cardiometabolic diseases, etc. Although it’s still early days for many medical large language models (LLMs), this state-of-the-art AI is starting to show promise, for example, by reducing the time it takes doctors to go through patient notes and helping with front-line decision-support—however, a lack of representative data leads to inequitable benefits (Davis N., Winawer N.). AI is also going to be key in improving coordination and reducing response time in disaster relief and rescue efforts, as they become more frequent with climate change. In the climate crisis, AI-based early warning systems for heatwaves, for example, are already in place in cities like Shanghai and Philadelphia, and others are emerging. Similarly for flood prediction, efforts by Google and others have led to flood prediction systems in places like India and Bangladesh to help plan for and respond to extreme monsoons. However, as with any technology, AI is meant to augment human/expert knowledge and not to replace it. This is particularly important in healthcare where people’s lives and health are involved in safety-critical situations.

Skenazy: How might AI exacerbate health inequity? Alternatively, how might it support solutions to limit harm and reduce inequitable impacts?

Khalid: AI depends on data, which means data can be the Achilles heel of AI. Let’s take the example of a clinical risk prediction AI. If there are biases in the underlying data or underlying algorithmic processes, the resulting AI tool can misestimate a real person’s health risks. At an individual level, this impacts a person’s health outcomes, and at a societal level, it can over- or under-prioritise entire population sub-groups. We have seen numerous examples of this, for example, when healthy black males were wrongly invited for COVID-19 screening, mothers-to-be of colour routinely have disproportionately been offered C-section procedures, and numerous other examples across the clinical spectrum. This is why it is important to input good data, and data that are representative. In the Ethnicity, Equity and AI projectfor example, for the first time, self-identified ethnicity data for over 60 million people has been completed in their electronic health records. The idea is that by using such representative data resources, health research can become less biased and scientists can build more representative health AI tools in future.

The Global Grand Challenges in AI has done some good work to showcase examples of where and how AI can be catalysed for reducing global health inequities. One such project that I have had an opportunity to contribute to looks at the issue of health data poverty and tests off-the-shelf LLMs for bias when applied to low- and middle-income country (LMIC) settings. We are starting to see that existing foundation models trained predominantly on data from traditionally data-rich countries in the Global North are biased but can be adapted to local settings in the Majority World by responsible and fair fine-tuning to local data. It’s important to keep in mind that data too has a carbon footprint. So the need for green AI cannot be overstated. Efforts by CCAI and others in the community towards this are also commendable.

Skenazy: Your work with the Planetary Health Informatics Lab is demonstrating how the mash-up of climate and health data can help address some of the challenges around equitable health for people and the planet. What are some of the challenges, and can you talk about any unique considerations for researchers when bringing together health data with other types of data—from satellite monitoring, environmental, etc.—in machine learning analyses? How might these considerations vary in different international contexts?

Khalid: The first thing to be said is that this is a young field, albeit a fast emerging one. The good news is that the compute power and high-performance infrastructure exist. Linking health and satellite data for example is complex, but with advanced data curation approaches it is achievable. Sometimes it is the seemingly straightforward issues that can be the most challenging. Let me give you one simple example—there is no universal definition of a heatwave–because what falls within and outside a “normal” temperature range differs from region to region. This may seem trivial, but getting it wrong can lead to biased models that may be appropriate for use in one region but not in others. So in many ways, it all boils down to good data. So here in the Planetary Health Informatics Lab we have a motto of “better data, better models, better health for all”.

Skenazy: Considering that climate and health impacts are and will likely continue to be the most drastic in the Global South, what are some of the main hurdles to using AI as a tool to increase health equity?

Khalid: The COVID-19 pandemic made it clear that “no one is safe until everyone is safe”. If you look at the latest IPCC reports the same can be said of the climate crisis – with evidence suggesting Europe and North America are warming at accelerated levels, while heatwaves in South Asia, South East Asia and Latin America continue to break records. Infectious diseases are spreading beyond traditional regions and borders, coupled with antimicrobial resistance. The main hurdle I see that touches all of these interrelated crises is a lack of cross-sector collaboration. We need to collaborate across our traditional in all aspects from data gaps, funding, partnerships with the public and private sector, and legislation, in order to ensure no one is left behind.

Additional reading and resources

For further reading, Professor Khalid recommends the following resources:

Planetary Health Informatics Lab website
PHI Oxford LinkedIn
@PHI_Oxford X Account
Predicting malaria outbreaks using earth observation measurements and spatiotemporal deep learning modelling: a South Asian case study from 2000 to 2017 Khalid S., The Lancet Planetary Health
2020 Report, Lancet Countdown
The Imperative for Climate Action to Protect Health Haines A. and Ebi K., The New England Journal of Medicine
Climate Change, Extreme Heat, and Health Bell M. and Gasparrini A., The New England Journal of Medicine

Unlocking Energy Flexibility of Residential Buildings

2024-07-01T00:00:00+00:00

In the shift towards decentralized and renewable dominated energy grids, residential buildings play a crucial role. Their flexible demand is key to integrating renewables like solar and wind, which are crucial for a sustainable low-carbon future. Building energy management relies on energy forecasting, improving flexible demand, and scheduling charging/discharging of batteries and EVs, all areas that can leverage artificial intelligence and machine learning in the short and medium term.

Energy flexibility¹ relies on energy users modulating or adapting their energy demand when needed, typically in response to a pricing signal or an emergency signal, e.g., to reduce air conditioning load on a hot afternoon. One under-utilized tool is continuous energy management, e.g., constantly adjusting loads as exemplified by smart home controllers.

The lack of a suitable benchmark environment has always been a bottleneck for developing such smart home controllers and highlighting their usefulness. Our CityLearn project is working to provide that benchmark environment.

Typically, smart home controllers focus on objectives like maximizing self-consumption or minimizing energy use or cost, with most projects designed as one-off developments. However, when buildings become part of the power grid infrastructure, integrating them becomes a multi-faceted and multi-objective problem, as their loads and potential PV generation, especially when aggregated, directly influence power generation needs. In addition, the design efforts are complicated by two ongoing challenges: integrating occupants and users into smart home energy controllers and simulation and understanding building loads as buildings increasingly electrify their current loads (e.g. space heating) and new loads are added at fast speeds (e.g. electric vehicle charging).

Here is where our CityLearn project has been at the forefront of development for the past several years. Instead of focusing on one particular aspect of demand-side management, we deliberately focused on integrating many distributed energy systems into a single computational environment to facilitate the development and benchmarking of advanced controllers. For example, we have shown that AI-based controllers can unlock energy flexibility by uncovering individualized occupant patterns and preferences, in a way that time-of-use pricing could not, leading to 10-15% reduced carbon emissions and >20% lower peak demand. In the same paper we found that a purely time-of-use controller did not reduce the aggregated peak at all, but merely shifted the peak load to a different time, which is unexpected at best. Similar unintended behaviors have been found by modulating smart thermostats, when all thermostats increase load at the same time due to their default setting.

Beyond these research findings, CityLearn has always been an open-source project; intended to be reused, updated, edited, and challenged by the community. We have held several versions of the CityLearn Challenge to crowdsource and accelerate innovation in this space. The most recent edition in 2023 has been made possible by the generous CCAI Innovation Grant.

In addition to the building control track which was included in previous years, we teamed up with Cambridge University’s Energy Efficient Cities Initiative which designed a forecasting track. The winners received a total of $5000 sponsored by Mitsubishi Electric Research Labs and presented their solutions at our NeurIPS 2023 Workshop. We have also created educational material, e.g., tutorials to get started quickly with CityLearn, for example at the CCAI Summer School. Most recently, we have integrated CityLearn into a course on Occupant-Centric Grid-Interactive Buildings.

The CCAI Innovation Grant 2023 also enables us to further push the capabilities of CityLearn, such as the development ofa methodology to integrate DOE’s recently released End-Use-Load-Profile database, which comprises representative models of all US buildings. This leap forward will allow us to estimate energy flexibility nationwide in the US. The methodology is generalizable to any location where archetypical building models exist and thus can be extended worldwide to facilitate the energy transition.

We are planning a major release of CityLearn in the coming weeks: in collaboration with Concordia University and the Politecnico di Torino, we are integrating occupant behavior models that can study how occupants interact with smart home controllers; and in collaboration with the Instituto Superior de Engenharia do Porto, we are integrating the much awaited EV module to study the impact that EVs can have on grid-aware demand side load management. These updates to the CityLearn platform move us closer to a vitally important, fine-grained computational environment to support emission reductions from buildings.

Thanks to CCAI’s generous support, we are getting closer to our vision for a more energy-flexible built environment, setting new benchmarks for researchers and practitioners.

The Climate Change AI Innovation Grants Program 2023 is supported by Quadrature Climate Foundation, Google DeepMind, Global Methane Hub, and the Canada Hub of Future Earth.

Notes

There are a lot of technical terms to describe energy flexibility which can almost be used interchangeably: demand side energy management, virtual power plants, load shedding/shifting, demand response, or demand dispatch ↩

Announcing the 2023 CCAI Innovation Grants Awardees

2024-06-26T00:00:00+00:00

Climate Change AI is excited to announce the awardees of our 2023 Innovation Grants Program. This program, now in its second cycle, supports impactful projects leveraging AI and machine learning to address problems in climate change mitigation, adaptation, and climate science, as well as the creation of publicly-available datasets/simulators to catalyze further work in this area.

To select our 2023 cohort, submissions from 47 countries were peer-reviewed by an international committee of experts in AI and climate change-relevant fields. The awardees include nine outstanding projects spanning 30 universities, companies, nonprofits, NGOs, and governmental organizations across nine countries and five continents, bringing together interdisciplinary teams for projects touching many aspects of the climate crisis. A key aspect of all the projects is a well-developed pathway to climate impact, generally instantiated through strong partnerships with entities key to deployment.

Funding awarded by primary proposal subject area.

The funded proposals are listed below (in alphabetical order), and more information is available at www.climatechange.ai/innovation_grants.

	Artificial intelligence for water management in the Red River Delta to meet the water demand and control saline intrusion in a changing climate Ivan Serina, Roberto Ranzi (Università di Brescia), Ngo Le An (Dai hoc Thuy Loi), Toan Trinh (University of California Davis) This project utilizes AI techniques to optimize water management in the Red River Delta in Vietnam, where a complex network of dams serves multiple purposes, including hydropower generation, flood regulation, agriculture water supply, and controlling seawater intrusion. The AI algorithms will offer optimal management policies for reservoir operations, focusing on water supply for agriculture and energy production during low-flow seasons while addressing constraints and climate conditions
	Curbing Illegal Logging Patterns using Sound-Based Detection Techniques Clara Zemp (Université de Neuchâtel), Henry Muchiri (Strathmore University), Anthony Mwangi (Kenya Forestry Research Institute), Fengshou Gu (University of Huddersfield) Illegal logging in Kenyan community forests is contributing to increased carbon emissions and negatively impacting the livelihoods of local communities. This project aims to use IoT devices to detect logging sounds, sending alerts for further investigation, and conducting carbon stock measurements to assess the impact, promoting data-driven forest management for climate change mitigation and benefits to local communities.
	Data Extraction and Modelling from Plant Trait Literature Richard Reeve, Christina Cobbold (University of Glasgow), Neil Brummitt, Ana Claudia Araujo, Ben Scott (London Natural History Museum), Claire Harris, Glenn Marion (Biomathematics and Statistics Scotland) The project aims to create a comprehensive global database of plant traits related to climate and habitat, utilizing computer vision and natural language processing techniques to process data from natural history collections and ecology literate. This database will provide essential information to improve global biodiversity-climate models, in turn contributing to policy efforts to preserve biodiversity in the face of changing climatic conditions.
	Developing machine learning tools to rapidly assess the catastrophic impact of a range-extending sea urchin in a global warming hotspot Arie Spyksma, Kelsey Miller, Katerina Taskova (University of Auckland), John Keane, Nicholas Perkins (University of Tasmania), Ariell Friedman (Greybits Engineering) In Australia and New Zealand, growing climate-induced sea urchin population growth poses a significant threat to kelp-dominated reefs. To address the threat, this project aims to develop machine learning to rapidly analyze underwater imagery to assess sea urchin expansion and kelp decline, providing essential data for proactive ecosystem management.
	EMPIRIC_AI: AI-enabled ensemble projections of cyclone risk for health infrastructure in Pacific Island Countries and Territories Christopher Horvat, Michelle McCrystall (University of Auckland), Berlin Kafoa (The Pacific Community), Elizabeth McLeod (World Health Organization), Craig McClain (Harvard Medical School) Communities in Pacific Island Countries and Territories (PICTs) face significant climate change impacts and their healthcare infrastructure is expected to be under extreme pressure in the coming decades. This project will develop high resolution climate models and targeted cyclone risk projections for health infrastructure in PICTs. This work will provide decision-makers and the public access to new datasets and tools with dedicated, high-resolution projections of future risks and planning to support health resilience for frontline communities.
	Flood Justice and Adaptation in the Rio Grande Valley of Texas with AI and satellite imagery Beth Tellman, Lucas Belury, Zhijie Zhang (University of Arizona), Ana Laurel (Texas RioGrande Legal Aid) Climate change is expected to increase flood risk in the Rio Grande Valley of Texas, as well as exacerbating inequities in exposure to floods, which disproportionately impact communities of color. However, inadequate flood risk data hampers adaptation efforts. This project plans to use machine learning and satellite imagery to create high-quality flood maps and a flood database for the Rio Grande Valley in order to support flood justice lawsuits and promote equitable flood adaptation.
	From Observing Power to Improving Power: Loss Localization in the Distribution Grid through Topology June Lukuyu (University of Washington), Genevieve Flaspohler, Mohini Bariya, Joshua Adkins, Kwame Abrokwah, Noah Klugman (nLine Inc.) Meeting the energy demands of low- and middle-income countries (LMICs) while addressing climate change requires more efficient electric grids. The project, based in Ghana, aims at a cost-effective solution involving voltage sensing at customer connections and machine learning algorithms to detect and localize distribution grid losses, particularly in the low-voltage networks where most issues occur.
	Mapping Rice Water Management and Methane Emissions in Ghana Sherrie Wang (MIT), Soren Vines, Freddie Monk (Aya Data), Benjamin Adevu, William Hunt (Demeter Ghana) Rice cultivation, while essential for agriculture, comes with methane emissions that are often poorly quantified in global climate analyses. This project aims to better understand and quantify methane emissions from rice cultivation, particularly in Ghana, by using satellite imagery and machine learning alongside field surveys to map rice flooding, a key factor influencing methane generation.
	The CityLearn Challenge 2023 Zoltan Nagy, Javad Mohammadi (UT Austin) Reducing energy consumption and emissions from buildings is crucial, given that buildings contribute to 30% of greenhouse gas emissions. This project develops a representative model of energy use in neighborhoods, and executes a novel edition of the CityLearn Challenge, which tests the ability of advanced control agents for demand response and reduction of carbon emissions at the neighborhood level. The model aims to reduce energy use and / or shift operations to low emissions electricity.

_{The icons used in the table above are obtained via remixicon.com.}

We would like to thank Quadrature Climate Foundation, Google DeepMind, and Global Methane Hub for their financial sponsorship of the 2023 Innovation Grants Program. We are also grateful to the Canada Hub of Future Earth for serving as the fiscal sponsor.

For those interested in applying for our Innovation Grants program, the next cycle of the program is expected to open in the coming weeks.

NeurIPS 2023 Workshop: Blending new and existing knowledge systems

2024-04-19T00:00:00+00:00

In December 2023, CCAI hosted a workshop at NeurIPS for the fifth consecutive year since 2019. NeurIPS is the largest conference in the AI field, today attracting thousands of participants across academia, industry, and government. The workshop brings together some of the latest research in AI that has potential to address the climate crisis. The 2023 iteration of the workshop focused on blending new and existing knowledge systems to inspire work that considers how novel machine learning research can build upon layers of institutional wisdom.

This year, 118 papers were accepted to the CCAI workshop, as well as 12 proposals and three tutorials. Of these, three papers were awarded prizes for “Best ML Innovation,” “Best Pathway to Impact,” and “Overall Best Paper.”

Overall Best Paper: Large Scale Masked Autoencoding for Reducing Label Requirements on SAR Data

In this paper, Allen et al. demonstrate a self-supervised approach to making satellite imagery (SAR data) classifiable during adverse weather conditions as well as night time hours. This advancement allows for remote intervention at rapid timescales during climate risk events. The novel breakthrough here was that the authors used SAR data (rather than typical RGB optical data), which can be detected at night and through adverse weather. SAR data is generated by satellites emitting microwaves down to Earth and then detecting the reflected waves to create a distinct image.

Figure 1 from Allen et al. that shows a representation of the raw SAR data, as well as the model predictions and ground truth labels. The model appears to predict the ground truth labels well.

The authors used a self-supervised pre-training regime with a masked autoencoder to create a base model. They then fine-tuned the pre-trained model on two downstream tasks: a regression task (MODIS Vegetation Continuous Fields), and a classification task (ESA World Cover). It was found that the pre-training significantly helped with the downstream tasks. This pre-training reduces the need for expensive hand-labeled data, thereby making SAR a more attractive data source for remote sensing applications. Added to the fact that SAR data can be used in all weather conditions and at night, this makes this work particularly exciting in the remote sensing field.

Best ML Innovation: Real-time Carbon Footprint Minimization in Sustainable Data Centers with Reinforcement Learning

Sarkar et al. present a novel real-time data center control agent that minimizes carbon footprint, energy consumption, and energy costs. This is in the context of the World’s ever growing use of data centers, and growing energy intensity of data centers. Additionally, the paper takes into consideration battery storage optimization, HVAC energy, and flexible load of data centers. The authors present the use of a multi-agent reinforcement learning (MARL) agent, that achieves a 10% performance improvement across emissions reductions, cost, and energy savings.

Figure 3 from Sarkar et al. showing the system map of the internal and external dependencies.

The authors plan to open source their DC-CFR framework to move the community forward in reducing carbon emissions of data centers. The reviewers particularly liked how clearly written the paper was, as well as the use of a novel method to target an important problem in the energy domain. Given the dramatic rise of data center usage (especially for AI training!), reducing the carbon intensity of their energy consumption will be key to helping us get to net zero.

Best Pathway to Impact: Discovering Effective Policies for Land-Use Planning

The global carbon balance is affected by both fossil fuel emissions and land emissions. These land emissions can vary widely depending on what the land is used for. For example, urban land typically releases more net carbon emissions than secondary forest vegetation (vegetation that has been touched by humans).

In this paper, the authors set out a framework for building a predictive model of land use change, that can then be used to evolve policies to find optimal policies. The authors combined data both from the land use harmonization project (LUH2) and from the sampling of the bookkeeping of land use emissions (BLUE) simulator. This training data set was used to build the surrogate model that was used by the prescriptor.

By searching through the solution space, the authors were able to construct a pareto front that trades off emissions reductions with land use change (the idea being that higher land use change might reduce economic impact). In doing so, policy makers can explore the optimal change, given a certain land use change tolerance they might have.

Demonstration of Land Use Planning tool, that the authors built.

The authors also built a useful exploration and visualization tool, to help policy makers navigate the solution space. The tool is available at https://landuse.evolution.ml, and was valued by the reviewers as an important component to connect the work to stakeholders.

Tips for great workshop papers

The reviewers were grateful for all the submissions they received at this year’s NeurIPS workshop. Having so many submissions demonstrated the growing world of applying machine learning to address climate issues, and inspired the team with hope for our future. Among the winning papers, a few things stood out that made them great papers. Namely, these were that the papers were easy to read, linked the work to potential impact in the climate space, and placed the work in the context of previous research done in the area. The team at CCAI looks forward to workshops at future machine learning conferences, and would strongly encourage anyone thinking about submitting to do so.

The Climate Change AI Summer School: 2023 Recap and 2024 Announcement

2024-03-28T00:00:00+00:00

At the end of August 2023, the Climate Change AI Summer School 2023 officially came to an end. The Summer School included both a selective in-person program held in Montreal and an open-to-all virtual program. The program provided resources and opportunities for participants with backgrounds in climate or machine learning related fields to tackle major problems at the intersection of climate and machine learning.

Over 10,000 registrants tuned in for the virtual program which taught participants about 11 different topics in the climate change and AI space. Through lectures and tutorials led by experts in these fields, participants gained insight and hands-on experience into the work being done in this domain. In-person participants worked together over the course of a week in Montreal to develop climate change solutions using AI and learn from top experts in the field.

In this blog post, we’ll share the details of the 2023 Summer School and how to participate in the 2024 Summer School.

Virtual Summer School 2023

The open-to-all virtual program spanned approximately 6 weeks and included lectures and tutorials on 11 different topic areas in the AI and climate change domain. Participants tuned in to the program from over 160 different countries. We are proud to serve a truly global audience. All of the lectures are publicly available on YouTube and have amassed nearly 60,000 views.

Experts from leading institutions like MIT, Columbia, Microsoft, Mila, Terra.do, ETH Zurich, and ESA Φ-Lab presented lectures and tutorials on AI for Transportation, AI for Climate Science, AI for Power Systems, AI for Agriculture, and more. One of the participants remarked that the depth and quality of these materials made the program “feel like a handbook” on applications of AI on climate. We are fortunate to have had such great instructors and tutorial developers contribute to the Summer School.

Along with the materials provided, the Summer School online community platform allowed for participants to interact with each other, asking questions and building relationships. Many participants were able to expand their network globally from the connections that they made during the program.

The 2024 Virtual Summer School will include additional materials and events. Participants will be challenged to think critically about some of the most pressing issues related to climate change and how machine learning can aid in finding solutions for these issues.

In-Person Summer School 2023

Interdisciplinary collaboration is key to developing solutions to climate problems using AI but it can often be challenging to work across disciplines effectively. Deploying those solutions into the real world and achieving measurable impact is even tougher. To help participants drive real change on climate change problems with AI, the CCAI In-Person Summer School focused on the key skills of collaboration, development, and deployment.

Participants had opportunities to collaborate virtually prior to the start of the Summer School, but the real work began in Montreal. In only a week, each team was able to develop, at a minimum, a proof of concept of their idea. The participants also attended learning exercises from experts in academia and industry related to communicating impact effectively, pitching, responsible AI, and evaluating a project.

“The Summer School was a turning point in my career. The experience, the inspiration, the opportunity provided and knowledge I acquired are indelible and eternal. It brought experts across-the-board globally in one place for idea generation and fertilization. I made new friends, networks, established collaboration and increased my social capital via the incredible people I met in the summer school.” - In-Person Summer School Participant

The teams worked on incredibly interesting projects and quickly put into practice the concepts that they learned throughout the week. On the final day of the Summer School, they put their ideas to the test by pitching their projects in front of a panel of experts.

The projects focused on topics like predicting flooding in coastal areas, optimizing Cassava yield in Nigeria, and forecasting land degradation in Colombia. Many of the teams are continuing to work together on their projects, with some earning acceptance of their papers and proposals at venues like the Tackling Climate Change with Machine Learning workshop at the Conference on Neural Information Systems Processing (NeurIPS). We can’t wait to see how these projects continue to progress!

We’d like to thank CIFAR, DENVR Dataworks, Mila, Volkswagen Group of America, and Worldsphere.ai for their support of the CCAI Summer School. Without them, the program would not be possible.

We are immensely grateful for all of those who participated in and contributed to this program. We are encouraged by the support that the Summer School received and will continue to build on this success in the future to further our goal of democratizing education in the climate change and AI space.

Below you will find more information about some of the 2023 in-person Summer School projects, mentors, and participants as well as how to participate in the 2024 Summer School.

Summer School 2024

We are excited to host the 3rd edition of the CCAI Summer School in 2024. As in 2023, the 2024 Summer School will include both an open-to-all virtual program and a selective in-person program. The virtual program will feature another stellar lineup of instructors and will cover additional topics that weren’t included in the 2023 program. There will also be many opportunities to collaborate with other participants and interact with instructors and tutorial developers.

The in-person program will bring a new cohort together to develop solutions to climate change-related issues using machine learning. If you want to enhance your ability to work in interdisciplinary teams at the intersection of climate change and AI, receive training from experts in this space, and develop global connections, apply here.

For more information about the 2024 Summer School and to register for the 2024 Virtual Summer School click here.

Teams and Team Members

Forecasting Land Degradation in Colombia

Wenxin Yang

PhD Student in Geography at Arizona State University

Wenxin is interested in the intersection of spatial data science and conservation. Her research centers on the effectiveness of protected areas in terms of monitoring, reporting, and predicting future trends, and informing conservation planning decision-making with spatial data science. Climate change has been impacting and will continue to impact protected areas efficacy whereas its trend and mechanism remains understudied and poorly integrated into conservation planning. Wenxin is interested in leveraging ML algorithms to promote forward and future looking conservation planning by developing novel datasets and models to start filling the research gap.

Tabea Stoeckel

Sustainability and Climate Change Consultant at PwC

Tabea has a background in environmental policy and how to use emerging technologies to fight climate change. After Tabea studied international affairs in Switzerland where she grew up, she worked for the Swiss government for 2 years in San Francisco on innovation policy, where she first worked on AI by organising a workshop around ethical AI. She then pursued a master’s in Environmental Technology at Imperial College London and started working for PwC’s Sustainability and Innovation team in London, where she co-authored several reports around e.g. the sustainable use of emerging technologies, i.e. blockchain, and how to direct funding to more impactful climate technology solutions.

Precipitation Intelligence

Matthias Bittner

PhD Candidate at the Christian Doppler Laboratory for Embedded Machine Learning at TU Wien

Curiosity and empathy have always been the driving forces in Matthias’s life.

During the journey from a student to a fully educated electrical engineer and ML scientist, his curiosity was strongly influenced by feeling the need to have a positive impact.

Right now, he is a PhD candidate at the Christian Doppler Laboratory for Embedded Machine Learning at TU Wien. There, he focuses on developing resource-efficient (low energy, latency, storage) machine learning time series applications, such as, forecasting grid loads, or precipitation downscaling of global climate models.

Sanaa Hobeichi

Postdoc at University of New South Wales (UNSW) Sydney

Sanaa works on applying Machine Learning to advance Climate Change research. Her current work is focused on developing Machine Learning methods for downscaling climate data and improving predictions of climate extremes. She is interested in explainable and physics-informed machine learning.

Sanaa obtained her PhD in Climate Science from UNSW, and she holds a MSc in Environmental Remote Sensing and a BSc in Computer Science and Applied Mathematics.

Ozioko Remigius Ikechukwu

Junior Lecturer at the University of Nigeria

Ozioko Remigius is from Nsukka in Enugu State, South East Nigeria. Nsukka is a beautiful soft-green rolling hilly City with very clement weather all through the year. It harbors the University of Nigeria Nsukka where he works as a Junior Lecturer. As a PhD student, his teaching and research revolves around Climate change governance and policy, food systems, community sustainability and engagement-nexus. Currently, he integrates AI and ML in developing solutions to the challenges of climate change to food production in Nigeria. His research team has deployed AI in the management, production and marketing of treasured vegetables in Nigeria.

Flooding Assessment for Coastal Environments (FACE)

Debasish Mishra

PhD Student in Biological and Agricultural Engineering at Texas A&M University

Debasish’s research focuses on studying the impact of extreme weather events on the soil-plant-atmosphere system using global satellite data. He aims to develop a data fusion model to predict flash floods and droughts, while also investigating the underlying scaling behaviors and controlling factors of these hydroclimatic extremes.

Tarini Bhatnagar

Solutions Architect at NVIDIA

A Geophysicist turned data professional; Tarini is a Solutions Architect at NVIDIA working on the frontiers of Artificial Intelligence. She leads technical customer engagements in Western Canada helping them adopt NVIDIA technology. She comes from a background in Earth and Environmental Science with extensive research experience and wanted to leverage my knowledge of both fields. Earlier such opportunities were limited, however with increasing relevance of Climate Change, new avenues have opened enabling her to contribute to this field. With so much Earth Observation data at hand, accelerated computing will help to achieve breakthroughs and I want to be a part of that effort.

Yusuf Ogunfolaji

MSc Student at the University of Ljubljana

Yusuf is from Nigeria and a second-year MSc student at the University of Ljubljana, Slovenia, studying Environmental Engineering.

During the first year of his double master’s degree at the University of Calabria in Italy, he assessed the impact of climate change on reservoir functionality using the HBV model to simulate the catchment runoff.

This experience triggered his interest in addressing this climate change impact using AI/ML as an alternative to the hydrological model.

Cassava Yield Optimization in Nigeria

Oluwaferanmi Oladepo

Undergraduate Student at the Federal University of Technology, Akure

Oladepo Oluwaferanmi is a Nigerian undergraduate at the Federal University of Technology, Akure, and a freelance software developer. As a member of the Oyo State Youth Parliament, he represents local youth, focusing on sustainable development and climate advocacy. This role exposed him to rural communities’ climate challenges and led him to discover AI’s potential in addressing them. He’s since become passionate about AI for climate solutions and has begun learning and applying AI techniques in this context, aiming to make a meaningful impact.

Gisa Murera

Senior Data Scientist at the National Bank of Rwanda, Kigali City

Gisa is a senior data scientist at the National Bank of Rwanda, Kigali City in the heart of Africa. He is passionate about harnessing AI and remote sensing for agricultural sustainability. Hailing from Rwanda, He’s been involved in data-driven research for years. His journey into climate and AI work began with a desire to address the challenges facing the agriculture sector and dynamic food prices. Currently, his focus is on leveraging the power of satellite imagery and seasonal agriculture surveys to develop optimal predictive machine-learning models for crop yield estimation and prediction. I am dedicated to using cutting-edge technology to drive sustainable agricultural practices in Rwanda and beyond.

Yazid S. Mikail

Data Scientist, Policy, Climate Change, and SDGs Advocate

Yazid has a professional certification in; Environmental Education and Community Engagement from Cornell University, Project Management Professional from the University of Virginia Darden Business School, and Leadership training from Harvard University. Yazid is from a semi-urban community named Dogarawa in Kaduna State Nigeria. Yazid had firsthand experience of the impact of climate change at 6 when his family’s house and farm were flooded like many other people in his community. Excited about the potential of technology in solving such problems inspired Yazid to link his firsthand experience to his book knowledge of data science, and climate change activism to explore the field of AI and ML for climate change.

WisePower Living

Donna Vakalis

Postdoc at Mila - Quebec Artificial Intelligence Institute

Donna worked in the fields of architecture and engineering before joining Mila, where she is currently a postdoc co-supervised by Dr. Yoshua Bengio and Dr. David Rolnick. Her research merges artificial intelligence with architectural engineering practices, toward mitigating emissions from new and existing buildings. She is also a former two-time Olympian and now the Scientific Lead for Racing to Zero, a non-profit that assists sport organizations to understand and reduce their negative environmental impact.

Elena Fillola Mayoral

PhD Student at the University of Bristol

Elena is a PhD student at the University of Bristol, working on using ML to accelerate greenhouse gas emissions monitoring. Her current research revolves around emulating atmospheric dispersion models, enabling more efficient simulations of the movement of gases like methane in the atmosphere. Elena enjoys working in interdisciplinary teams, using her background in Machine Learning to solve problems across a wide range of areas. In particular, she’s interested in Earth Sciences and in projects that have real-world impact.

Alina Klerings

PhD Candidate at University of Mannheim

Alina is from Germany and holds a Master’s degree in Data Science. Her growing concern with climate change got her interested in the intersection of machine learning and climate related applications. Over the past years she was involved in various forecasting projects in the energy domain, including grid load and net loss prediction as well energy disaggregation. She has industry experience through her work at a transmission system operator and this fall she started her PhD in AI Systems.

High Impact Climate and Weather Events Detection

Arbel Yaniv

Electrical engineering Ph.D. candidate at Tel Aviv University, Israel

Equipped with extensive background in algorithms and machine learning implementations & research in the energy field including power flow, distribution systems optimal operation, photovoltaic systems, optimal storage control and non-intrusive load monitoring (NILM).

Constanza Molina

PhD Student at the University of Munster

Constanza is from Chile and completed her master’s degree at the Pontifical University of Chile, where she worked with physics-informed neural networks to model global surface dust deposition during the Holocene and Last Glacial Maximum periods. Currently, she is pursuing her PhD at the University of Munster in Germany, and is part of the WilDrone project. This program is dedicated to wildlife conservation through the utilization of autonomous drone technology and AI for monitoring wildlife populations, tracking their movements, and managing human-wildlife conflicts.

Rosa Pietroiusti

PhD candidate at the Vrije Universiteit Brussel, Belgium and the University of Warwick, UK.

Rosa is investigating how climate science research is being used as evidence in child and youth-led litigation around the world. She is focusing on fields including impact attribution, extreme event attribution and intergenerational analysis of global climate and impact models.

Benedetta Mussati

Student in the AIMS DPhil program at the University of Oxford

Benedetta is an Italian student in the AIMS DPhil program at the University of Oxford, researching continual meta-learning algorithms.

She sees machine learning as a very helpful tool for climatologists to get a richer and clearer understanding of the data available and make more informed decisions. ### Vegetation Forecasting for Efficient Carbon Markets

Dr. Sylvia Smullin

Head of R&D at VEIR

Dr. Sylvia Smullin is an experimental physicist with more than a decade of experience at the early stage of research and development. Motivated by the magnitude and urgency of the challenge, Dr. Smullin chose a career in industry so as to work on impactful climate solutions. Currently the Head of R&D at VEIR (veir.com), she has worked at both bigger companies (PARC, Google X) and at startups (Makani Power, Form Energy). Collaborations at Alphabet sparked her curiosity about and appreciation for AI tools. Across technologies, Dr. Smullin brings a passion for collaboration, data analysis, and impact-focused work.

Dr. Mike Smith

Director of AI at Aspia Space

Dr. Mike Smith is originally from Hatfield – a small town on the outskirts of London – and has a background in astronomy and physics, with his masters concentrating on using deep learning to diagnose heart disease and his PhD concentrating on applying deep generative models to extract information from crazily huge (petabyte-scale!) astronomical survey data. Now he’s working as the Director of AI at a geospatial start-up that spun out of his PhD research. At Aspia Space they use deep learning to extract useful information from remote sensing imagery (like foundational modelling for drought and flood prediction, crop identification, and cloud cover removal from satellite imagery).

Ricardo Barros Lourenço

PhD Student at McMaster University

Ricardo is a Brazilian Geoscientist pursuing his PhD at McMaster University in Canada, where he works as a Research Assistant in Geography. His academic journey has been a rich tapestry—starting in geology, moving through a master’s in data science, and advancing into computational science for remote sensing. He also lived in Chicago for almost five years, deeply involved in research using generative models for spatial tasks. Currently, his passion lies in crafting explainable models for vegetation dynamics under various climate change scenarios. Methodologically, his work leans on time series analysis, ensemble learning, and neural networks to make spatial computing more understandable and impactful.

Sustainability Report Chatbot (Susi)

David Denny

Co-Founder and Partner at Carpe Diem Developers

David has worked in the solar industry for the past twelve years. He has founded and led multiple companies that collectively have installed several thousands solar panel systems. He is committed to powering the renewable revolution, and sees AI as a tool to lower the cost and increase the pace of our clean energy transition.

Flora Haberkorn

Data Scientist at the Federal Reserve Board

Flora Haberkorn is currently working as a Data Scientist for the Federal Reserve Board (FRB) in the division of International Finance. She specializes in big data engineering and AI/ML techniques. She typically works with LLMs across a variety of domains such as climate, trade, and macroeconomics. Her research interests focus on the application of artificial intelligence for economic policy and research projects. She became a Data Scientist after pursuing the use of advanced analytics to enhance ongoing news and social media analysis as a Research Assistant at the FRB.

Sara Badran

MENA Regional Coordinator at Thought For Food

Sara Badran, holding a Bachelor’s degree in Food Science and Management and a Master’s in Environmental Sciences, is dedicated to leveraging her expertise to drive innovation and foster sustainability, particularly in addressing climate change. Serving as the MENA Regional Coordinator at Thought For Food, Sara passionately supports innovative and entrepreneurial endeavors. Her experience includes diverse projects in waste reduction, and food supply chain optimization through participation in accelerators. Sara’s research extends to applying machine learning models in her thesis, focusing on the circular economy’s potential in waste management and reuse, showcasing her commitment to sustainable solutions.

Mentors

Sharon Xu

Senior Data Scientist at Indigo Ag

Sharon is a senior data scientist at Indigo Ag, where she is designing deep learning models to detect regenerative practices and inform carbon credit generation. In previous roles, she has led spatial modeling R&D for field experimentation, designed and productionized scalable ML models at Tripadvisor, proposed novel methods to model human dynamics at MIT, and prototyped graph CNNs to identify security incidents at Element AI.

Arthur Ouaknine

Postdoctoral Researcher Fellow at McGill University and Mila - Quebec Artificial Intelligence Institute

After completing his PhD in the automotive industry in collaboration between Institut Polytechnique de Paris and valeo.ai, he wanted to use artificial intelligence to better understand the biodiversity and climate crises. His projects are now focused on computer vision and deep learning applied to forest monitoring. He also a core team member of Climate Change AI leading the webinar team.

Millie Chapman

Postdoc Fellow at the National Center for Ecological Analysis and Synthesis

Millie recently completed her PhD at University of California Berkeley in Environmental Science, Policy, and Management and is currently Postdoc Fellow at the National Center for Ecological Analysis and Synthesis. Her research is at the intersection of decision theory, ecology, and data justice, asking how we can leverage AI to devise more effective and equitable strategies to meet global biodiversity targets under uncertainty.

Christopher Yeh

PhD student at the California Institute of Technology

Christopher Yeh is a 4th-year PhD student in the Computing and Mathematical Sciences (CMS) program at the California Institute of Technology (Caltech). Advised by Professors Yisong Yue and Adam Wierman, Chris’s research combines theoretical tools from uncertainty quantification and online decision-making with applications in sustainable energy systems. He is a Caltech Resnick Sustainability Institute Scholar and is funded by a Caltech Amazon AWS AI for Science Fellowship. He received B.S. and M.S. degrees in computer science from Stanford University where he worked with Professor Stefano Ermon, and he also spent a year in Beijing as a Schwarzman Scholar at Tsinghua University.

Using Machine Learning to Forecast the Weather and Climate

2024-02-07T00:00:00+00:00

Intro

Climate change has enormous implications for extreme events and hazardous weather.
ML offers unprecedented potential to predict such events and thus adapt to and mitigate their effects.
Our three forecasting tutorials illustrate end-to-end pipelines that use ML tools to predict extremes and climate variability.

As I flew across the eastern coast of Canada in mid-June 2023, it was impossible to overlook the hazy smoke clouds caused by the ongoing wildfires. As a climate scientist, it’s hard to ignore the link between these fires and the massive carbon emissions we humans release into the atmosphere each year - a staggering 37+ billion tons of CO2! While this gas remains invisible to the human eye, the cumulative impact of depositing such enormous quantities of it into the atmosphere doesn’t; across time and space, climate change is now materializing into phenomena such as the rapid spread of flames, heat waves, air pollution, and even alterations to planetary-scale climate modulators such as El Niño.

Using forecasting to adapt to the present weather conditions is a vitally important undertaking. However, it’s also difficult due to the weather’s diverse spatial and temporal nature. Take, for example, wildfires. They are difficult to predict and model as it is hard to get accurate input data for such models. Wildfire models involve tiny spatial scales and require pin-point accurate atmospheric data (such as wind, which would transport the fire) and fuel measurement.

Traditionally, for wildfires, or similar short-range extreme weather events, we use a predictive framework tailored to short time scales. However for climate we use a different prediction framework to capture the long-term time scales and extensive spatial coverage. Yet weather and climate are intrinsically linked and influence each other.

Efforts to use a modeling framework where both weather and cliamte are represented in one system started in the ’90s with the idea of “seamless prediction” (Shuka, 1998). Seamless prediction uses a single framework for forecasting across a range of timescales from days to years and represents one of the ultimate aims for weather and climate prediction. Realizing that aim requires finer and finer spatial and temporal resolutions calculated over long time periods, which translates into a prodigious computational load. This presents a problem as while our computers have gotten bigger, they haven’t necessarily gotten faster. With the limits of current computation, our climate models are suited to short-term, high-resolution simulations. Hundreds of years of 1km or less spatial resolution is impossible without substantial computational costs.

Machine Learning (ML) methodologies and accelerated computing harnessing GPU technology offer an unprecedented potential to provide faster, accurate, long-term predictions as an alternative or a complement to traditional, model-based methods (more about this here).

Our series of forecasting tutorials illustrate how to produce such fast and accurate predictions for three different types of phenomena spanning different temporal and spatial scales: (1) El Niño, (2) synoptic scale weather, and (3) extreme events at the edge of synoptic scale.

The ML methodology behind all three tutorials is Convolutional Neural Networks (CNNs). Three spatial and temporal scales are addressed with one methodology suggesting that we are moving closer to a seamless prediction framework.

Forecasting the El Niño Southern Oscillation (ENSO) with Machine Learning

ML prediction of ENSO (orange) versus ‘ground truth’ ENSO (blue). Plot produced as part of the tutorial.

The El Niño prediction tutorial walks the users through a hierarchy of ML-based predictions on an indicator of ENSO from simple (regression-based model) to complex (deep learning, CCN-based) models. Dynamical models forecast El Niño by modeling the physics of the atmosphere and ocean (Cane et al. 1986) while ML methodologies mostly use data produced by long climate model simulations, and in some cases observations for the prediction. CNNs have successfully predicted El Niño with superior skill to the dynamical models which are typically used (Ham et al. 2019). The ENSO tutorial walks the student through multiple statistical tools to predict ENSO with discussion questions that invite the student to think and adapt the tool to their own need.

One of the current challenges in ML is how to constrain deep learning approaches such as CNN - which are based on learning from large amounts of data - by incorporating the laws of physics into the model. Incorporating physics can provide models with structure to facilitate learning as well as provide interpretability of performance. The El Niño prediction tutorial raises this problem along with pointing out the importance of the physical interpretability of the model, prediction, and results. Physical interpretability e.g. understanding what our model is doing under the hood, how the model is making its predictions, and the physical considerations to make a certain prediction (and not make it differently) all represent a cutting-edge research area for both climate science and ML.

FourCastNet: A practical introduction to a state-of-the-art deep learning global weather emulator from Pathak et al., 2022

FourCastNet is a state-of-the-art deep learning-based surrogate for weather models, a tool that can be used to produce predictions without the computational cost of running the full weather model. FourCastNet is a CNN trained on ERA5, a reanalysis dataset consisting of hourly estimates for several atmospheric variables at a latitude and longitude resolution of 0.25 degrees and at elevations from the surface of the earth to roughly 100 km covering from January 1940 to the present.The model generates a week-long 25km resolution forecast in less than 2 seconds, orders of magnitude faster than global forecast systems (GFS). Beyond the day-to-day weather, FourCastNet can capture several different types of extreme events, such as hurricanes. FourCastNet has been used successfully to predict a very broad range of extreme events,for example Africa’s hottest recorded heatwave in July 2018 as well as the 2023 hurricane season.

Accurate, reliable, and efficient means of forecasting global weather patterns are of paramount importance to our ability to mitigate and adapt to climate change. While the training of the model is made a priori and theCNN weights are provided directly in the tutorial, the users are shown the evaluation pipeline to understand how they might train such a model.

FourCastNet is trained using significant computational resources (both in terms of input data and GPU time). This tutorial gives a rare view into such a model as by making the trained model weights, code, and datasets freely available. The availability of FourCastNet aids new research, for example by allowing the creation of rapid and inexpensive large-ensemble forecasts to complement and augment numerical weather predictions (NWP) based forecasts.

FourCastNet prediction versus ‘ground truth’ as represented by ERA5 data. Plot produced as part of the tutorial.

While FourCastNet performs rapid forecasts, there are a number of limitations. As with many ML-based methodologies, FourCastNet does not perform data assimilation with observations, relying on ERA5 to provide real-time initial conditions. The model is also a fully data-driven model without physics constraints. Additionally, the current resolution of 25 km is too coarse to capture finer scale structures of sub-grid processes. Lastly, FourCastNet is intended as a weather model and its behavior on climate time scales is not understood well. The authors however envision a coupling between a climate model output and FourCastNet in order to understand extreme weather in future climate scenarios for climate range timescales.

ClimateLearn: Machine Learning for Predicting Weather and Climate Extremes from Nguyen et al. 2023

The extreme temperatures and precipitation in the summer of 2023 have made clear the impact that extreme weather can have. Fundamental to climate change mitigation and adaptation is (1) understanding the changing likelihood of extreme conditions in long-term observational records and (2) exploring the frequency and intensity with which such extremes might occur, especially at regional scales.

The last tutorial in the series, “Machine learning for climate extremes”, has two components: (1) a prognostic component where extremes are forecasted at a coarse resolution and (2) a mapping of the coarser resolution to a finer one. Downscaling – mapping coarse resolution data to finer resolutions - aims to bridge the gap between large-scale, dynamical models and the end users who need localized forecasts and projections. Using ML, the tutorial combines temporal forecasting and statistical downscaling, providing a model which makes reasonably accurate predictions in most parts of the globe, except the polar regions.

The ML prediction (temperature in Kelvin) and the difference between the prediction and the ground truth (left). Plot produced as part of the tutorial.

The visualization shows that the model makes reasonably accurate predictions in most parts of the globe, and the prediction well correlates with the ground truth. The model seems to make large errors near the two poles, where the temperature is more unpredictable.

Conclusions and Summary

Through our series of Colab-based tutorials, you will get the chance to learn hands-on how to use a set of ML-based models and large amounts of data to make informed weather or climate-related predictions within hours. While ML methodologies in atmospheric science have a long way to go, it is essential to note that these fast, accurate, and reliable predictions were almost impossible to envision even a few years ago, let alone be available to anyone with a good Wi-Fi connection. Moreover, such predictive tools go beyond the proof of concept phase and can be used operationally (e.g., FourCastNet).

From a scientific perspective, our tutorials offer the potential for the users to make not only predictions but also pursue climate-informed projects; each of these tutorials in itself can play a threefold role:

they can be an end-to-end data analysis project spanning different temporal and spatial phenomena in weather or climate, while still providing a foundation for forecasting other phenomena;
all three tutorials represent machine learning-based tools ready to be used operationally and tuned for various applications the user might have in mind; and
all these three tutorials can be seen as a tool to tackle deep fundamental scientific questions in climate and weather prediction including, for example, tuning such tools for a seamless prediction framework or new phenomena (e.g. wildfire prediction).

As a complete data analysis pipeline, our projects start with the first APIs for accessing popular repositories that host global climate data and walk you through all the analysis steps, ending with visualization packages for plotting the results. The projects also lead the students through ML architecture considerations, for example, hyperparameters tunning for each type of phenomenon, looking into the right optimization metrics for the right problem, evaluating model performance, and much more.

Ultimately, these three tutorials represent a multi-faceted view and a hands-on way of exploring the rapidly changing boundaries of forecasting the behavior of a changing climate through AI. We increasingly face climate manifestations such as heatwaves, extreme El Niños, or forest fires, tools like the ones described in these tutorials and the unprecedented amount of data we possess are an important part of adapting and thriving to our changing climate.

Cane, M. A., Zebiak, S. E. and Dolan, S. C. (1986). Experimental forecasts of El Niño. Nature, 115(10), 2262-2278.

Ham, Yoo-Geun, Jeong-Hwan Kim, and Jing-Jia Luo. “Deep learning for multi-year ENSO forecasts.” Nature 573.7775 (2019): 568-572.

Shukla, J., 1998: Predictability in the midst of chaos: A scientific basis for climate forecasting. Science, 282, 728-731.

Using Machine Learning to Increase Durability and Reduce Returns for Sports and Fashion Goods

2024-01-27T00:00:00+00:00

Fashion shapes the way we express ourselves and reflects personal values. However, from haute couture (handmade) to fast fashion, the cost of fashion extends beyond the price tag. This article delves into the far-reaching consequences of fashion products and how overlooked factors such as quality can be combined to significantly reduce their impact on the planet.

A product that breaks quickly or fails to meet consumer expectations has a staggering impact. Durability issues give rise to short lifecycles, resource squandering, and increased landfill waste. Poor designs create unwanted products. Sizing mismatches generate unnecessary emissions through returns. The prevalence of poor quality not only undermines sustainable consumption but also accelerates environmental damage. A report from the European Parliament quantifies the existing impacts of textile production on the environment.

The fashion industry accounts for 4 percent of global GHG emissions, according to McKinsey. According to the World Bank, every year the fashion industry uses 93 billion cubic meters of water; about 20 percent of wastewater worldwide comes from fabric dyeing and treatment, and 87 percent of the fabric is incinerated or disposed of in a landfill. Overall, the textile industry contributes about 10% to global emissions, which, at the current pace, may rise to 50% (according to the World Bank) by 2030.

With the aim of reducing the impact of our industry, a team of data scientists and machine learning engineers gathered together to discover the information which can improve product durability and reduce returns using available data.

Product Reviews to Measure Product Life Cycle and Quality

In this context, we looked at different data sources to gather information on the quality and durability of the product, focusing on aspects that matter most to consumers and capture circularity performance. Circularity metrics at the product level were not available to us, and to our knowledge, there are still many challenges with assessing holistic footprint and circularity performance.

Since we did not have circularity metrics, we focused on proxies such as durability and quality-led returns (understood as the match of consumer expectations and product claims). Durability is key, as the most sustainable product is the one that is not produced, and if we produce products that last two times longer, we are directly reducing our footprint by 50%. Given that many returned products are incinerated or piling up in warehouses, we also sought data sources that could help us understand the reasons for product returns so we can minimize them.

Our available data to capture information about prodcut returns included product test data, expert reviews, and product reviews. Product test data provides valuable insights into the internal testing done on a product under lab conditions; however, the problem is that it is not systematically stored and does not give us visibility into the reasons for returns. Expert reviews provide feedback on the product across all relevant dimensions (versatility, durability, fit, etc.) but rarely cover most products (only key franchises), and it does not necessarily match what actual consumers report about the products. Our preferred option was, therefore, the product reviews from actual consumers of the products, as it gives us a good understanding of the positive and negative features of the product, as well as the reasons for return or a short life cycle. In our analysis, we were able to cover more than 80% of the revenue of the products we want to analyze with English language product reviews, and expanding the language scope would lead to more than 95% product coverage. For this reason, we chose product reviews as the data source to measure product aspect sentiment.

This post proposes a novel approach using public product review data. By using Natural Language Processing, consumer feedback can be extracted, analyzed, and grouped by key topics or aspects to gain remarkable insights.

How Aspect-Based Sentiment Analysis Can Improve Product Design and Consumption Decisions

In particular, we will use Aspect-Based Sentiment Analysis (ABSA) to extract consumer sentiments from reviews. The framework leverages pretrained or fine-tuned language models to detect word relationships in the text, which is later used to map aspects and sentiments. We were interested in analysing dimensions(from now on referred to as ‘aspects’) that truly matter both for product footprint and consumer satisfaction.

Although ABSA is able to find most of the aspects available in the reviews without customization, we choose to limit the aspects to a predefined list due to the diversity of consumer experiences with a product. Based on product expertise, six key categories, or quality aspects, were defined. They cover most of the topics reported by consumers found in our data.

The Quality Aspects are:

Fit
Design
Material
Comfort
Durability
Color

A score is calculated for each of the above aspects. This score is calculated by analyzing the sentiment associated with every predefined keyword using the pyabsa (Python Aspect-Based Sentiment Analysis) framework. These sentiments are then amalgamated to create a score for each aspect. With this information, we can detect products that have been underperforming and understand why. These underperforming products need immediate attention by either fixing the underlying problems or improving the product in the next release.

The spider plot below gives a visual comparison of two products using the Quality Aspects.

The Spider plot allows to compare and understand the performance of multiple products

In this example, the product in red is superior in almost every dimension, particularly fit and color, while not substantially different in the others. One can also see that the blue product design cycle should focus on fit and color problems.

These scores for each metric can be used in a number of ways to tackle environmental challenges.

Fit:

Helps by reducing returns and minimizing the need for shipping back and forth due to sizing issues, decreasing the carbon footprint associated with transportation and packaging materials.

Accurate sizing reduces the need for extra production and inventory, which, in turn, lowers manufacturing emissions and resource consumption.

Design:

Designing products that have timeless appeal and can withstand changing trends reduces the frequency of product turnover, resulting in fewer items becoming obsolete and contributing to landfill waste.

Longer-lasting designs mean fewer resources are used for constant redesign and production, ultimately reducing emissions linked to manufacturing processes.

Material:

Material score helps with choosing durable and sustainable materials that can extend the product’s lifespan, reducing the need for frequent replacements.

Sustainable material replacements can also have a smaller footprint during production, as they might require less energy or fewer chemicals in their manufacturing processes.

Comfort:

Comfortable products lead to longer usage and reduced turnover. Consumers are less likely to discard a comfortable item, decreasing the frequency of purchases and waste.

Longer product lifespans mean fewer products need to be manufactured, cutting down on resource consumption and associated emissions.

Color:

Offering classic colors that stand the test of time can reduce the “fast fashion” cycle, where colors fall out of style quickly, leading to frequent discards.

Fewer color variations and shorter design cycles can lessen the strain on manufacturing facilities and reduce emissions related to production.

Durability:

Producing durable products that withstand wear and tear means consumers won’t need to replace them as often, reducing the overall demand for new products.

Durable products require fewer replacements, which lowers the demand for manufacturing new items and the associated emissions from production processes.

These metrics provide invaluable insights into product quality, offering opportunities to reduce environmental harm, enhance future product development, and curtail the overproduction of unwanted items. When combined with other performance indicators like return rates, these insights not only contribute to environmental sustainability but also bolster business efficiency.

Our analysis of these quality metrics revealed the following key findings:

39% of products achieved a fit score below 80%
6% of products received a fit score below 50%
2% of products scored below 50% in material quality.
80% of products received a material score exceeding 80%.

This shows, in our sample, that there is significant room for improvement in fit, while material is normally well perceived by consumers. There are further insights that can be obtained for a particular product as well as the entire range. These results are shared to illustrate how this data can drive quality and design strategy.

Our Results, Limitations, and Training Resources

This methodology isn’t without its shortcomings and drawbacks. One of the biggest drawbacks is that by the time enough review data is collected, the damage process has already started. We recommend leveraging expert reviews as a proxy when insufficient consumer data is available. Another challenge is that, in many cases, products with issues have already been produced in large quantities due to large lead times and the limited producer capacity. If insights from consumer reviews are coupled with Make to Order strategies and localized manufacturing, it will reduce inventory requirements while increasing agility in the feedback cycle and the potential to enhance quality improvements.

In conclusion, using aspect-based sentiment analysis on consumer reviews offers a promising avenue to increase consumer satisfaction thereby reducing returns and excess inventory, ultimately leading to decreases in the carbon footprints of fashion goods. By harnessing the power of consumer feedback and NLP, we have the potential to not only create better, longer-lasting products but also to mitigate the environmental toll of these products and pave the way for a more sustainable and responsible future.

If you are interested in leveraging this methodology for your research or industry, we have published a notebook with Kaggle data that walks you through an analysis using ABSA.

Our mission

We are a passionate group of data scientists and machine learning engineers committed to using AI to reduce product footprint and improve product decision-making.

From left to right: Murtuza Kazmi (top), Ali Naeem (bottom), Anis Mazika (middle), Alan Fortuny (top), Afolabi Lagunju(bottom)

Introducing The ForestBench Project

2023-11-28T00:00:00+00:00

In the realm of environmental science, accurate data is not just a luxury; it’s a necessity. This is particularly true when it comes to understanding the carbon content stored in the world’s forests—a critical factor in the global fight against climate change. Forests are the lungs of the Earth, absorbing carbon dioxide and releasing oxygen, thereby playing a pivotal role in climate regulation. However, existing data sets for estimating forest carbon content have been largely skewed towards the Global North. This bias leaves a glaring gap in our understanding of forests in the Global South, which are often rich in biodiversity and play a unique and pivotal role in carbon sequestration.

We here present the ForestBench project, which is designed to address this imbalance by gathering high-quality, “gold-standard” data from forests that have been historically underrepresented in global research. In this blog post, we introduce our community-driven approach, and take you on a tour around the world to our project sites. We want to share our many lessons learned here, both the technological and the cultural ones.

Methodological Innovation: A Dual Approach to Data Collection

What sets ForestBench apart is its comprehensive approach to data collection, which combines traditional ground-level measurements with state-of-the-art aerial drone technology. On the ground, the project involves the meticulous identification of tree species and the measurement of tree diameters—factors that are crucial for accurate carbon content estimation. Above the canopy, camera drones capture high-resolution data, providing a detailed layer of information.

To succeed in our combined approach, we included experts from many backgrounds in the project, working across disciplines to achieve a holistic data collection. For the drone imagery, we worked closely with local technology companies that offer drone-related services. The ground-level data collection on the other hand needed to be tuned towards each specific ecosystem, and so we invited local organizations that work on biodiversity, conservation, and restoration to collaborate with us. With these local experts on board, we could obtain unique data that takes the circumstances of each project site into account.

Community Engagement: Beyond Data Collection

While the collected scientific data are the core of ForestBench, what truly elevates the project is its commitment to community engagement and social responsibility. The initiative actively collaborates with local communities of Indigenous Peoples, who are the most knowledgeable about their local ecosystems but are rarely included in scientific research. ForestBench not only provides financial compensation for community participation but also emphasizes the transfer of knowledge and skills, thereby fostering a model of inclusive and sustainable scientific research.

Exploring the Conservation and Restoration of Blue Carbon Habitats: A Dive into Mangrove Forests

In recent times, the importance of conserving and restoring our planet’s natural habitats has gained significant attention. Among these habitats, blue carbon ecosystems, particularly mangrove forests, have emerged as crucial players in the fight against climate change.

Oceanus Conservation is a non-profit organization that has dedicated its efforts towards the conservation and restoration of blue carbon habitats. Camille Rivera, the local project lead, along with her team, has been instrumental in collecting ground-level data. Their excellent work has been instrumental in collecting data on about 3,000 mangrove trees, including species identification and Diameter at Breast Height (DBH) of trees, ground-level photographs, and drone imagery. On top of that, the data is updated quarterly, ensuring that the information remains current and relevant to monitor the development of their conservation sites, such as the Cagwait Mangrove. This is pivotal in understanding the potential of future reforestation projects in mangrove forests, and would not have been possible without Camille and her team. From our perspective, this collaboration was the most fruitful and productive aspect of our project.

Collaboration with Oceanus Conservation to collect mangrove forest data in the Philippines. Top: Our locations at Cagwait, with the types of trees to be encountered in these pristine mangrove forests. Bottom: Data collection, measuring Diameter at Breast Height (DBH) of a mangrove tree, and locations of some trees identified near the field station. Images by Camille Rivera.

Adventures in Paraguay: A Journey of Collaboration and Conservation

The heart of South America, Paraguay, is a land of contrasts. From its bustling urban centers to the serene landscapes of the Chaco, it holds a myriad of fascinating places. Our collaboration with the Ministry of Environment of Paraguay unveiled some of these mysteries, offering us a unique perspective on the nation’s commitment to conservation.

A Warm Welcome from the Ministry

Our journey began with an invitation from the Ministry of Environment and Sustainable Development (MADES Paraguay) in February 2022. Eager to establish a relationship and facilitate local data collection for our project, we found ourselves in the company of passionate individuals dedicated to preserving Paraguay’s rich biodiversity. The meeting with the Minister and his team was enlightening. It became evident that Paraguay was keen on supporting our data collection efforts, especially in the face of an impending climate crisis that threatens the nation with heatwaves and drought.

Into the Wild: The Defensores del Chaco National Park

To our delight, the Ministry organized a field trip to the Defensores del Chaco National Park, the country’s largest national park. Within the larger Chaco region, this park is home to an old-growth semi-arid closed-canopy shrubland forest. This unique ecosystem, however, is under threat from rapid deforestation driven by agriculture and cattle farming.

Our visit to the park was both educational and experiential. We tested our drones and wildlife camera traps at the fringes of the pristine dry forest. The aerial photography provided invaluable above-canopy data, while ground-level data collection offered insights into the local biodiverisity. However, the slender and shrubby nature of the trees in this ecosystem presented challenges in measuring their properties.

Field-test data collected during our field trip to the dry forest of the Defensores del Chaco National Park, Paraguay. Top left, and bottom: Orthomosaic and 3D model of the ranger station at Fortin Madrejon, which oversees the activities of the Ministry of the Environment for maintaining the national park. Top right: One of our camera traps caught a Capybara searching for water at a nearby lake in the evening hours.

A Lifelong Guardian: Don Silvino Gonzalez

One of the highlights of our trip was meeting Don Silvino Gonzalez, Paraguay’s most senior park ranger. A guardian of the forests, Don Silvino has dedicated his life to their protection. His insights into the natural history of the area were invaluable. Moreover, we were fortunate to use some of the land he owns near the park as one of our data collection sites for the ForestBench project.

Hope for Progress

Our adventures in Paraguay were a testament to the power of collaboration. The passion and commitment of the Ministry, the wisdom of Don Silvino Gonzalez, and the beauty of the Chaco have left an indelible mark on our hearts. As we move forward, we carry with us the lessons learned and the hope for a brighter, greener future for Paraguay. Since our initial visit to the Chaco, we have collected more aerial data from Don Silvino’s land, and continue to support the region through our project.

Brazil: Guardians of the Amazon - The Kayapo Indigenous People

Deep within the heart of the Amazon Rainforest, the Kayapo Indigenous People stand as guardians of over 9 million hectares of primary forest. This vast territory, equivalent in size to Portugal, is a biodiversity hotspot, home to numerous endangered species. The Kayapo, with their profound connection to the land, actively combat large-scale deforestation and external threats. Their efforts are not just about preserving an ecosystem; it’s about safeguarding their ancestral culture and way of life. While the Brazilian government’s regulations have limited the collection of forest carbon data, the Kayapo have generously provided wildlife imagery, offering a rare glimpse into the rich tapestry of life in the Amazon.

Data collection with the Kayapo People in Brazil. Images by John Meisner.

Bhutan: Planting a Million Dreams

In the serene landscapes of Bhutan, a monumental initiative is taking root. The Million Trees Project, supported by the Bhutan Ecological Society, has set forth an ambitious goal: to plant one million trees over five years. With an average target of 200 to 300 trees per acre, the focus is on high-value tree crops that can provide sustainable income for rural farmers through fruit production and other materials. Beyond the economic benefits, this project is a beacon of hope for biodiversity conservation in Bhutan, acting as a buffer against the impacts of climate change. The data from this project, including GPS coordinates and species information for almost 4,000 young trees, provides invaluable insights into reforestation efforts in the region and was collected at two locations.

Future Endeavors and Impact

The ForestBench initiative is a pioneering project that seeks to comprehensively gather data on the carbon content stored within forests, particularly in the Global South. All collected data will be published open-source, paving the way for future work to focus on its analysis. This data can be instrumental in our understanding of forest carbon stock, especially in the Global South. The project is more than just a scientific initiative; it is a paradigm shift in how we approach environmental research, community engagement, and climate action. By leveraging cutting-edge technology and fostering meaningful community partnerships, ForestBench is setting new benchmarks (pun intended) in the accurate measurement of forest carbon content. Its comprehensive and adaptable methodology has broad implications for climate science, policy-making, and sustainable development. We encourage you to visit the project website for more information.

Deep learning of nanoporous materials for chemical separations

2023-10-08T00:00:00+00:00

Separations are foundational processes in the chemical industry, accounting for about half of US industrial energy use and more than 10% of the world’s total energy consumption. An analysis of the largest energy consuming industries indicate that replacing traditional separation processes with more efficient alternatives could potentially eliminate 100 million tonnes of carbon emissions and save billions of dollars in energy costs annually. Separation also serves as a cornerstone of carbon capture and storage, enabling the selective removal of carbon dioxide from pre- and post-combustion gas at power plants, as well as from the environment using emerging direct air capture technologies.

Separation using nanoporous materials — which function as molecular sieves to allow certain molecules to pass through while rejecting others (see Figure 1) — can be an order of magnitude more efficient than heat-driven distillation processes (i.e., boiling liquids to obtain separate fractions at different temperatures). However, designing nanoporous materials involves identifying the optimal material structure for a target application, which can be challenging due to the large number of possible materials to choose from.

Among the porous materials that could be used for separation, zeolites stand out as robust and industrially proven. To date, hundreds of thousands of potential zeolite structures have been computationally predicted. However, so far, only about 250 have been synthesized experimentally and even fewer are used for practical applications. Despite their great promise, testing even the hundreds of currently synthesizable structures would require an impractical amount of effort and resources.

Illustration of a tubular membrane separation module. Image taken from J. Membr. Sci. 520, 434-449 (2016).

In this project, a deep learning framework named ZeoNet was developed to rapidly and accurately predict the separation performance of zeolite materials for large, flexible molecules (see Figure 2). ZeoNet borrows ideas from the computer vision community, viewing materials as 3-dimensional images. On a standard GPU, ZeoNet can process more than eight structures per second. This speed is in stark contrast to state-of-the-art computer simulations that can take hours per structure and lab experiments that can take at least days. ZeoNet also generalizes to novel structures not seen in the training set.

The ZeoNet framework for predicting the separation performance of nanoporous zeolites, showing an example of producing ingredients for making high-quality lubricant oils.

Preliminary follow-up work also shows that ZeoNet, once trained on one system (i.e., to predict adsorption of a specific molecule), can be easily adapted to work for another application with significantly less training data (with as few as ~500 training samples). This suggests that the ZeoNet model captures a rich, universal representation of the 3-dimensional structure of zeolites. What’s particularly exciting is that this representation appears to allow highly specific discrimination about which computer predicted structures might actually be synthesizable — a grand challenge in zeolite science. Collaborations with experimental groups are currently underway to test these ideas in physical labs.

This work is published as a featured article in the Journal of Materials Chemistry A and we are also releasing the dataset and software code in the hope of encouraging more AI researchers to work on physical sciences and engineering problems.

The Team

ZeoNet was developed in collaboration between the College of Engineering led by PI Peng Bai and his students Yachan Liu and Samuel Hoover, as well as Prof. Wei Fan and the College of Information and Computer Sciences led by PI Subhransu Maji and his students Gustavo Perez, Aaron Sun, and Zezhou Cheng at University of Massachusetts, Amherst.

Mapping Species From Crowdsourced Data Using Machine Learning

2023-09-05T00:00:00+00:00

The users of community science platforms such as iNaturalist (www.inaturalist.org) generate millions of photographic observations each month documenting where different plant and animal species can be found. In the last few years, advances in AI in the form of automated image classifiers allow non-experts to identify the different species that are present in these images. However, automatic species identification in images remains a challenging problem, as community science platforms can potentially contain images from hundreds of thousands of different species. One of the major sources of difficulty is that there can be multiple species that look visually similar (e.g. crows versus ravens), and thus remain very challenging for AI systems, and humans, to visually disambiguate.

Knowing where a species of interest is likely to be observed (i.e. its geographical range) is a valuable piece of information when trying to determine what species might be present in an image. Inspired by this simple observation, in recent work, we have developed efficient machine learning models that can take the location where an image was captured as input to determine which species are likely to be present at that geographical location. These models are trained on the locations of observations from iNaturalist, though there are many other platforms that it can be applied to.

Predictions about which species occur at a specific location can be combined with the predictions from automated image classifiers to enhance the accuracy of the final species estimate. We have found that this approach improves species classification performance in images from 75% to 82% when evaluated on a dataset that contains images from 40,000 different species from around the world. Importantly, this improvement does not necessitate any changes in the underlying image classifier being used.

Does the image in the top left, taken in Central Europe, contain a European Toad or a Spiny Toad? By looking at the past history of such observations on iNaturalist, shown in red below, we can see that it is very unlikely for Spiny Toads to be found outside of Portugal, Spain, or France.

Motivated by this success, we next wanted to understand how effective these models trained on crowdsourced location data are at predicting the ranges of species on a global scale. Unfortunately, unlike in the image classification example above where we could easily measure the improvement in the final image classification performance, assessing the accuracy of a predicted range for a species is a much more challenging problem. This is because, with some exceptions, there are no established benchmarks for this task.

To address this problem, we developed a suite of spatial prediction tasks for evaluating the performance of machine learning-based species range estimation models. Our new benchmark will enable researchers to evaluate different, and complementary, spatial prediction tasks, from estimating species’ ranges to quantifying how well these models help resolve mistakes in image classifiers. Despite being trained on spatially biased and incomplete data, we were able to show that it is possible to obtain performance that is within 75-80% agreement of expert-derived ranges across a wide set of species. These are promising initial findings, but also point to the fact that there is more work to be done to further improve performance on these spatial prediction tasks.

Example predictions from one of our machine learning-based species range estimation models. Here we illustrate the results for four different species, where brighter colors indicate locations where the model is more confident that the species occurs. For reference, we add the map of current iNaturalist observations in the bottom right of each.

Main findings:

Once trained, our range estimation models are compact and efficient to run. A single model smaller than 50MB can be used to predict the ranges for 47,000 different species.
The performance of our models improves as we add more data during training, including from new species.
Models trained without any explicit environmental information beyond geographical location can learn effective representations of space and also embed species with similar ranges close to each other (see below).
Despite many of the biases present in the crowdsourced data we used to train our models, we observe that in many cases the predicted ranges are similar to the expert-defined ones.
How a model is trained can have a significant impact on its performance. In the paper, we explore different ways of training these models and find that some ways lead to better performance. This opens many interesting doors for future progress on this task from both machine learning and statistics.

Our machine learning approach jointly encodes the ranges of thousands of different species into one compact model. On the left we visualize the learned representations for 47,000 different species in two dimensions, and on the right we can see the corresponding estimated ranges for a subset of them denoted by the red dot on the left.

Going forward there are many interesting open questions that require additional investigation. For example, species observations from community science platforms can exhibit large geographical and temporal biases as some locations can be difficult for community scientists to reach. In addition, there are global biases as some of the most biodiverse locations on the planet are not yet well represented on these platforms. Some species are also much easier to photograph and identify than others; thus, detectability and data imbalance needs to be explicitly factored into our models. Care needs to be taken so that potential mistakes made by the range estimation models do not adversely impact the image classifiers and result in inaccurate identifications being made. These open questions mean that great caution should be exercised if attempting to make any conservation decisions based on the outputs of these current models.

In the future, we can expect to see more species observation data being uploaded to platforms such as iNaturalist. As the global reach of these online communities grows, we will also potentially obtain more data from currently underreported, but highly biodiverse, regions. Addressing the open methodological issues highlighted above will lead to a new set of methods that will be able to use this data to reliably estimate the ranges of previously mapped species. The resulting improved understanding of how different species are spatially distributed around the world will be an important piece of information in our quest to better conserve them.

More information about this work can be found in our recent paper titled “Spatial Implicit Neural Representations for Global-Scale Species Mapping” (https://arxiv.org/abs/2306.02564). The paper includes a detailed description of the models and our benchmark evaluation datasets and also includes a detailed discussion of the results. The work was presented at the International Conference on Machine Learning in July 2023.

Team members:

Using Machine Learning to Integrate Mangrove Restoration with Sustainable Aquaculture Intensification

2023-07-21T00:00:00+00:00

Contributors: JC Nacpil, JT Miclat, Oshean Garonita, Anica Araneta, Joseph Schmidt, Rod Braun, Jack Kittinger, Pia Faustino, and Dane Klinger

Shrimp aquaculture has grown 100-fold over the last 40 years, from an estimated 74,000 metric tons in 1980 to 7.4 million metric tons in 2020. This rapid growth has come at the cost of critical coastal ecosystems, especially mangroves. While deforestation rates have decreased from 0.21% (1996-2010) to 0.04% (2010 to 2020), at least 35% of global mangroves were deforested in the late twentieth century, and the ecosystem services they provided remain lost.

Developed by Conservation International (CI), the Climate Smart Shrimp (CSS) program supports communities’ livelihoods and food security while also improving coastal resilience and adaptation to climate change. The initiative provides the resources for small- and medium-sized farmers to sustainably intensify production on a portion of their farm in exchange for mangrove restoration on the remainder of the farm. This enables smaller farms to be more competitive with the global commodity market while providing sustained funding and opening available parcels for coastal mangrove restoration. But not all aquaculture farms are suitable for CSS as an approach.

In partnership with Thinking Machines, Arizona State University, and Konservasi Indonesia, this project used machine learning and earth observation data, such as openly available Planet NICFI satellite imagery and Clark Labs aquaculture pond data, to identify and classify aquaculture farms in Indonesia and the Philippines that use extensive (as opposed to high-productivity or intensive) production methods. The team then combined this information with open data on sea level rise, flood risk, infrastructure access, historical mangrove cover, and other attributes to identify viable sites for CSS. Identifying a pipeline of these optimal sites accelerates CI’s ability to engage farmers, industry, and communities, and attract investment to scale CSS.

The project’s main output is a web-map tool that analyzes the potential suitability of aquaculture sites according to preferred site characteristics. These site characteristics are separated into filtering and scoring criteria based on our defined attributes and applied to tiles identified as aquaculture areas. Each tile or tile cluster on the map must pass all the filtering criteria to be considered ‘suitable’ or is automatically considered ‘unsuitable’ if a tile or cluster fails any criterion.

Gif of web-map tool. For a demo of the web-map tool, contact data-for-development@thinkingmachin.es.

The interactive web-map tool was designed to streamline the implementation of CSS but has the benefit of helping inform and guide conservation practitioners in making decisions on where to focus other nature-based solution approaches. This tool makes it easier to identify areas that are both suitable candidates for restoring mangroves to increase forest cover and also viable for intensifying shrimp aquaculture to contribute towards food security and support local livelihoods. While the tool in its current form helps CI to rapidly evaluate the hundreds of thousands of potential hectares where CSS might be implemented and evaluate optimal locations, only slight updates or changes to the scoring criteria could make this tool applicable in a wide range of coastal and terrestrial restoration applications.

Developing this site assessment tool has rapidly accelerated the ability to identify and evaluate potential CSS sites across Indonesia and the Philippines and builds upon a larger portfolio of CSS projects. In addition to pilot projects on the ground to validate environmental, social, and economic indicators at the farm level, CI has been designing a dedicated Climate Smart Shrimp Fund as a revolving loan facility to pioneer ways of financially supporting the widespread implementation of CSS across geographies. In addition to priority sites in Southeast Asia, CSS is being piloted in Ecuador, one of the top five global shrimp producers, further demonstrating its application under various production systems, management, and geographies.

The site assessment tool, developed with support from the Innovation Grants Program, enables CI and its project partners to more efficiently and effectively apply CSS in shrimp aquaculture geographies to support livelihoods and food security while providing climate adaptation and resilience benefits for coastal communities.

Project Partners:

This project leveraged expertise from across academia, conservation organizations, and the technology industry.

Arizona State University (ASU) is a public research university in Phoenix, Arizona which houses the Julie Ann Wrigley Global Futures Laboratory. ASU’s sustainability work centers on education, research, business practices, global partnerships, and interdisciplinary solution initiatives.

Conservation International (CI) is a global non-profit organization that protects nature through fieldwork, working with Indigenous communities, governments, and corporations, and innovations in science, policy, and finance. CI’s priorities include climate stabilization through nature-based solutions, protecting the ocean, and expanding planet-positive economies.

Konservasi Indonesia (KI) is a national foundation that aims to support the sustainable development and protection of critical ecosystems in Indonesia. KI is the main implementing partner for CI in Indonesia and works with governments and other stakeholders to design and deliver innovative nature-based solutions for climate change and protect terrestrial, coastal, and marine ecosystems for the benefit of people and nature.

Thinking Machines (TM) is a technology consultancy that builds cloud artificial intelligence and data platforms to solve high-impact problems for large organizations across Southeast Asia. With at least a third of TM’s projects driven by the company’s mission to use data for good, TM has worked with diverse partners to collaborate on projects where data science can support and strengthen solutions for climate action and social impact. TM is also a portfolio company of the UNICEF Innovation Fund and has presented original research at top machine learning conferences including NeurIPS and ICML.

Last year, the CSS team wrote a blog post on their expected outcomes for this project. You can check it out here.

Using Reinforcement Learning to Improve Energy Management for Grid-Interactive Buildings

2023-06-02T00:00:00+00:00

Buildings consume a significant amount of global energy and contribute to greenhouse gas emissions (around 19% in 2010), but also have the potential to reduce their carbon footprint by 50-90%. Optimal building decarbonization requires electrification of end-uses and integration of renewable energy systems. This integration requires aligning availability of renewable energy with the energy demand, and must be carefully managed during operation to ensure reliability and stability of the grid.

Demand response (DR) is an energy-management strategy that allows consumers and prosumers to provide grid flexibility by reducing energy consumption, shifting energy use, or generating and storing energy. Buildings and communities that offer such two-way interaction with the grid are often called grid-interactive buildings or communities.

Advanced control systems can automate the operation of energy systems, but effective DR requires intelligent load control. Advanced control algorithms such as Model Predictive Control (MPC) and Reinforcement Learning (RL) have been proposed for such applications. RL is a type of machine learning that involves learning through trial and error and can take advantage of real-time and historical data to provide adaptive DR capabilities.

A major challenge for RL in DR is to compare algorithm performance, which requires a shared collection of representative environments to systematically compare building optimization algorithms. More specifically, in the context of building and HVAC (heating, ventilation, and air conditioning) systems, there are nine challenges that need to be addressed in order to make those environments practical for real-world applications. Inspired by Durlac-Arnold et al. we have summarized them in a recent paper as follows:

Being able to learn on live systems from limited samples
Dealing with unknown and potentially large delays in the system actuators, sensors or feedback
Learning and acting in high-dimensional state and action spaces
Reasoning about system constraints that should never or rarely be violated
Interacting with systems that are partially observable
Learning from multiple, or poorly specified, objective functions
Being able to provide actions quickly, especially for systems with low latencies
Training off-line from fixed logs of an external policy#
Providing system operators with explainable policies

In this post we show an example for Challenge 8.

Training off-line from fixed logs of an external policy

Challenge 8 is particularly important because in buildings there is typically a rule-based controller (RBC) available. An RBC is a deterministic policy, in essence a sequence of if-then-else rules that the controller executes. If historical data of the RBC policy (fixed logs of an external policy) is available, it could be used to jump-start a model-free RL controller, thereby reducing its learning time.

To investigate this challenge, we benchmarked some RL configurations in CityLearn, an OpenAI Gym environment for energy management in grid-interactive smart communities, using the dataset of the CityLearn Challenge 2021. This data consists of the energy usage of one medium office, one fast-food restaurant, one retail, one strip mall retail, and five medium multi-family homes in Austin TX. Data for a four-year period is available. We compared independent RL agents (using Soft-Actor-Critic or SAC) and coordinating RL agents (MARLISA algorithm) to a industry best practice (default settings) and an optimized (based on grid search) RBC . The agents have been pre-trained for one month (744 hours), six months (4344 hrs) or one year (8760 hrs) with either the basic or the optimized RBC. During the training period the building is operated with the respective RBC, and after the training period the RL algorithms are switched on. We then compare their performance across a set of key performance indicators (KPIs) in the following figure:

We show the result of each agent (MARLISA or SAC) normalized (taking as a reference the performance of the basic or the optimized RBC). One can observe that longer training periods or more historic data do not necessarily improve the results as all curves eventually end up very close together in each experiment. This suggests that little training data from a basic RBC is sufficient to jump start RL controllers. Another relevant insight is that both RL agents (SAC and MARLISA) improve the performance of the basic RBC (curves drop below 1), which is expected. However, if the RBC is sufficiently calibrated, i.e., the optimized RBC, both RL agents are not outperforming the RBC (curves go above 1 or baseline). This shows that the quality of the RBC or the available offline policy significantly impacts the results, more so than the available data itself.

Another challenge not listed before is the lack of a common benchmark environment to evaluate SOTA (state of the art) models. In our context, we can use standardized environments such as BOPTEST, ACTB, CityLearn or Sinergym. A common benchmark environment could trigger the same excitement and progress in the field as ImageNet did for deep learning. If you’re interested in learning more, there is an online tutorial to go along with slides presented at ICLR as well as a new community authored paper.

As we have shown in the blogpost, the use of RL can be a very powerful tool, particularly for suboptimal RBCs, without large training times and data sets. However, for optimized RBCs RL does not seem to provide benefits. Standardized environments like CityLearn overcome two of the main challenges of the field, which is the need for offline training data and common environments to define SOTA models in the field of RL like we have for computer vision.

To make the most of this, researchers in both building engineering and computer science should work together to transfer theoretical findings and new ideas into practical solutions for building energy management.

CCAI Core Team Profile: Priya Donti

2023-03-06T00:00:00+00:00

This interview has been edited for clarity and length.

Tell us about yourself.

I’m a co-founder and the Executive Director of Climate Change AI (CCAI), currently funded by the Runway Startup Postdoc Program at Cornell Tech. I’ll start as an Assistant Professor at MIT in Fall 2023.

Recently, I finished my Ph.D. at Carnegie Mellon in Computer Science & Public Policy, working on machine learning to dynamically optimize power grids to better integrate renewable energy. Before that I traveled for a year as a Watson Fellow to interview people about next-generation power systems and did my undergrad at Harvey Mudd.

How do you feel experiences like the Watson Fellowship have interacted with your current academic research?

I went into my undergrad at Harvey Mudd knowing I wanted to work on climate change, but it wasn’t clear to me how I would do that. I initially thought I’d be a materials scientist and develop next-generation solar panels, but I ended up enjoying my computer science classes much more than any others. This created a conundrum, because it wasn’t clear to me then how computer science could play a role in climate change.

I ended up taking a leap of faith — doing a computer science and math major, and pursuing environmental analysis through a minor and my extracurricular activities — and kept looking for ways to bring these interests together.

I eventually found a paper arguing that AI will be a critical component of next-generation power grids that are able to incorporate large amounts of renewable energy.

I got really excited about that topic and applied for the Watson Fellowship to learn more. During my Watson, I conducted interviews in five different countries to better understand what people meant when they used the terms “next-generation power grids” or “smart grids,” as well as what the technical, policy, and social considerations were behind modernizing the grid. I wanted to understand how these differed between countries and contexts. That exploration gave me a lot of perspective as I continued on to do my Ph.D.

Let’s talk about the leap of faith you mentioned. You knew you enjoyed computer science and you knew you wanted to work on climate change, but you didn’t quite see the intersection. You’ve done a great job of building the space that you wanted to inhabit. Do you have any reflections on that, or advice for people starting out?

Addressing climate change is going to take a really wide set of skills, tools, and approaches, and everybody has something that they can contribute. At the same time, it’s important not to approach problems with a hammer looking for a nail — different problems will require different skillsets, and even when a particular skillset is relevant, the way you actually use that skillset often fundamentally needs to change based on the realities of the problems you’re addressing.

For example, many machine learning tools have been developed primarily with large-scale image and text data in mind, and those methods don’t work out-of-the-box when analyzing (e.g.) power grid physics. This means machine learning needs to be approached differently in power grids than in some other settings. Setting up the necessary “feedback loop” between your problems and your skillsets requires actually doing your homework on the climate change areas you care about.

Making your journey publicly available can be a way to keep yourself accountable, and to provide a lot of value to other people as well. Take the “Tackling Climate Change with Machine Learning” paper that launched CCAI. We came in not with the conclusion that AI is necessarily applicable to climate action, but with a hypothesis. We looked through the literature, and we talked to people to understand whether that hypothesis was substantiated, and in what ways. We did not talk about applications where we didn’t think that AI was relevant.

I agree that the feedback loop is important. It’s also hard sometimes. What has helped you build that kind of feedback loop to make sure your work is interacting with the domain you’re working in?

The first step is to pick a domain. What are you interested in and excited about? Pick a domain that reflects that and then do a deep dive - read the literature, understand who the players are, go to the relevant venues where people are discussing these topics, follow people on social media, and generally get integrated into the community.

Many people try to over-optimize this choice of domain to try to find the “objectively best” one, and that decision paralysis prevents them from ever diving in. In reality, we need a diversity of people working on a diversity of problems across sectors - so it’s okay to be guided by your interest, plus a rough understanding of the order-of-magnitude of impact.

What are you excited about right now in CCAI?

I’m really excited by the impact that CCAI has had so far. We’ve brought together thousands of people through our conferences and events, including our inaugural CCAI Summer School in August. We’ve run multi-million dollar grants programs to fund impactful work, provided actionable policy advice through venues such as the Global Partnership on AI, and inspired the creation of many other initiatives through our work. What is sometimes unbelievable is that this work has largely been done by volunteers.

While we’re excited by the impact we’ve been able to have as a predominantly volunteer team, we’re ready to further scale our activities and our professional staff to further unlock the potential of AI for climate action. We’re currently looking for funding to hire staff, scale our existing activities, and launch new programs aimed at further catalyzing impactful work by bringing together the right information, organizations, and people. In particular, we want to foster better pathways to deployment for AI-for-climate work to make sure it has actual impact on the ground.

Working on climate change can be tough. What gives you hope or motivates you in the work?

To me, action is a very empowering thing. The climate is changing. The realities of that are really difficult, but we have the ability to work on creating a better future.

Are there specific people that are inspiring in this space?

Carla Gomes, the founder of the computational sustainability movement, has been a huge inspiration and mentor to me. She’s done a lot to show both that sustainability is an inherently important topic to work on, and that sustainability challenges can push the boundaries of computer science innovation in interesting and novel ways.

What would your superpower be and why?

Teleportation - I would love to be able to connect more easily with family, friends, colleagues, and others from around the world, as well as more easily travel to new places. Or, you know, maybe the ability to suck carbon dioxide out of the atmosphere and turn it into basalt - but perhaps that’s a little bit too on theme.

Are there any important lessons that you’d like to share, either from your academic path or from helping start this organization?

One is the importance of being humble and open to feedback. It can be tempting to sprint as fast as you can with the tools you have. But climate change problems are systems problems; the way you work on them interacts with considerations from many different sectors and fields. Being open to being told you might be thinking about something incorrectly and using that to strengthen your thinking is extremely important.

Relatedly, it’s important to collaborate, meet people, and expand your horizons to learn from those who were trained in different ways than you and who bring in different perspectives from their sectors, geographies, and lived experiences.

Using AI-driven Yield Estimation to Improve Resilience of Malian Cotton Farmers

2023-02-22T00:00:00+00:00

Climate Change Risks to Agri-Business Sector

15 million Malians rely on agriculture for food and income, much of which is small-scale agriculture. Among agricultural crops, cotton is one of the most important economic crops in Mali. Cotton contributes 15% of GDP and 11% of total value of exports (second only to gold), with cotton production supporting close to a quarter of Mali’s population, 4 million.

Cotton is also highly sensitive to climate change, as yields are closely linked to rainfall. Mali’s cotton production is at high risk due to the expected shorter growing periods, increasing duration of extreme temperatures, more frequent and intense droughts and rainfalls due to the climate change. Cotton production has already been affected by drought in 2012-13 and heavy rains in 2017-18. Moreover, Mali scores high in vulnerability and low in readiness against climate change, according to their ND-GAIN index, which summarizes a country’s vulnerability to climate change and other global challenges in combination with its readiness to improve resilience.

Addressing these challenges in order to ensure the sustainability and viability of Mali’s agri-business sector is of highest importance and provides an excellent opportunity for the use of advanced AI technologies. However, these technologies require accurate field data, which is scarce in underdeveloped countries. In these uncharted lands, there is no reliable source of information to help farmers manage their crop production risks.

The IMPRESSYIELD project addresses the lack of data to use AI to help farmers better understand and adapt to these challenges. This project introduces the dynamic yield estimation map, a tool that uses freely available earth observation data and advanced AI techniques to monitor climate change risks in Mali’s cotton production.

The Dynamic Yield Estimation Map

We propose an innovative yield estimation method at commune scale and field scale for cotton in Mali. We employ domain adaptation techniques, which enable operation with limited amounts of in-situ data.

The “Improving Resiliency of Malian Farmers with Yield Estimation: IMPRESSYIELD” project combines state of the art remote sensing AI and freely available satellite earth observations. There are several reasons why satellite imagery is useful for yield estimation in Africa:

Wide coverage: Satellite imagery can cover large areas quickly and relatively cheaply.
Repeatability: Satellite imagery can be collected at regular intervals, allowing longitudinal monitoring of crop growth and yield.
Accessibility: Satellite imagery is often more readily available than other forms of data, such as ground-based measurements, which are difficult to obtain in some parts of Africa.
Ease of use: There is dedicated software for satellite imagery that makes it easy to extract useful information from the data.
Non-invasiveness: Satellite imagery does not require physically accessing fields, which can be important in areas with restricted field access.

Overall, satellite imagery provides a cost-effective and efficient way to estimate crop yields in Africa and a valuable tool for agricultural planning and decision-making.

Despite the advantages of satellite imagery, satellite imagery is not widely supported by in-situ data from field visits which is required for validation and reliable dataset generation. Thus, state-of-the-art methods and functions which can be successfully employed in the regions with in-situ data access may be challenging or impossible to implement in Africa.

As a result, earth observation driven state-of-the-art crop analytics on monitoring climate change risks affecting the African agri-business sector are underutilized.

IMPRESSYIELD addresses the monitoring of climate change risks of cotton production by providing a specific AI-based tool, the dynamic yield estimation map, integrating freely available earth observation data and specific crop yield models with an advanced AI approach that achieves high accuracy with limited in-situ data. This map enables continuous monitoring of predicted yield from application regions, using a deep neural network based yield estimation algorithm based on the timeseries data generated from multiple satellite sources.

The Team

Field research team Samet Cetin (left) and Osman Baytaroglu (right) after a rainy day across the fields. Credit: Osman Baytaroglu;

The IMPRESSYIELD project is driven by a collaborative research partnership between Istanbul Technical University (ITU), Middle East Technical University (METU), and global Agritech startups Agcurate and OKO.

ITU team consists of 3 people, Esra Erten (principal investigator) along with 2 researchers Mustafa Serkan Isik, a PhD graduate in Geosciences and Mehmet Furkan Celik, PhD candidate in Geosciences in ITU. METU team consists of 2 people, Gokberk Cinbis, who is an assistant professor in METU along with Samet Cetin, who is a PhD candidate in Computer Science in METU, and who also works part-time in Agcurate. Agcurate team consists of 6 people. 3 of them (Berk Ulker, Burcu Suslu and Can Kayabek) are responsible for the development of the machine learning algorithms, 2 of them (Metin Emenullahi and Burak Isilar) are responsible for development, deployment and maintenance of multi-modal data processing infrastructure, and Osman Baytaroglu is responsible for overall project coordination and serving as a data quality engineer, and supporting the team via field visits (such as in Mali). Last but not least, Simon Schwall from OKO Finance coordinates OKO’s roles and responsibilities while Abdallahi Ould Mohamed, country manager of OKO Finance in Mali, is responsible for dissemination activities that are being organized in Mali.

Conclusion

The dynamic yield estimation map is a valuable tool for improving the resiliency of Malian farmers by using AI to monitor and adapt to the risks posed by climate change. By using freely available earth observation data and advanced AI techniques, the map provides farmers with valuable insights into the state of their crops and helps them make informed decisions about how to best protect their livelihoods. The IMPRESSYIELD project represents an important step forward in the use of digital technologies to support small-scale agriculture in developing countries, and has the potential to greatly improve the sustainability and viability of Mali’s agri-business sector and to provide risk management products, such as insurance, allowing farmers to be protected against the risk of climate variability.

This project was funded by the Climate Change AI Innovation Grants program, hosted by Climate Change AI with the additional support of Canada Hub of Future Earth.

Circularity in Fashion, powered by AI

2023-02-13T00:00:00+00:00

We are regularly reminded of the impact of our food, transport, and energy systems on our biodiversity and climate. However, fashion has enormous environmental impacts which must be addressed to mitigate climate change.

The United Nations Environmental Programme and the Ellen Macarthur Foundation show that:

10% of annual global carbon emissions come from the fashion industry
87% of the total fiber input is incinerated or disposed of in a landfill
20% of wastewater worldwide comes from fabric dyeing and treatment
93 billion cubic meters of water are used yearly
50 billion plastic bottles equivalent in microplastics is released yearly
Consumers buy 60 percent more than they did in 2000, and keep it half as long
The average garment is worn 10 times before disposal

Despite the current environmental impact, there are ways to address these issues. According to Mckinsey, 70% of the fashion industry’s emissions come from upstream activities such as materials production, preparation and processing which could be mitigated cheaply.

Artificial intelligence (AI) has been integrated into the fashion industry to help drive sales, with most use cases covering dynamic pricing and excess stock clearing, product recommendations to increase conversion, process automation, and trend detection.The following reviews found on Medium[1] [2] support such claims. Despite the potential for cheap abatement and the importance of the industry in global warming, AI hasn’t been used to reduce emissions and adopt circularity principles.

In this post, I provide case studies for the use of AI to accelerate the circular economy in the fashion industry, refocusing AI as a powerful tool for climate change mitigation, not just a sales tool. I also share my personal opinion as an AI developer in the fashion industry.

What can AI do to accelerate a truly circular fashion?

In the following, I provide a long but not exhaustive list of use cases where AI can be used to make the industry more sustainable in all stages of value creation, including the potential role of consumer behavior change and policy changes supported by AI.

AI can improve Design and Raw Materials Selection

1) AI can be used to design long lasting garments. We can use product reviews and lab data to predict how long products last, supporting the creation of those with the largest life cycles.

2) AI can be used to create truly unique articles leveraging computer vision, natural language processing, and other techniques to reduce assortment size while staying relevant for the consumers.

3) Instead of taking a cost based approach to design and pricing, we can use estimated perceived value and willingness to pay methods. This can support the creation of valuable and profitable products [1] [2] [3].

4) Leverage Reinforcement Learning and other methods to create designs that are easy to repair, recycle or decompose.

AI can help to eliminating excess stock and waste

1) AI can be used to reduce excess stocks via better demand forecasting and stock optimization.

2) AI can support pre-order models with zero stock and agile lead times.

3) AI can make renting and second life channel recommendations better.

4) Computer vision can be used to detect the status of garments for resale and second hand channels.

AI can be used to improve the delivery operations and transport emissions

1) AI can be used to reduce returns by making accurate product recommendations or verifying that the product fits before purchase.

2) AI can be used to find similar replacements that are in the same distribution center so that missing stock does not need to be flown in.

3) Using Green Last Mile AI power logistics to deliver products with renewable energy.

AI can help understand consumer behavior and design appropriate policies

1) AI can help to understand product design or policy interventions that lead to longer use of garments.

2) AI can provide insights into the real and perceived values that drive sustainability clothing choices.

3) AI can test theoretically appealing policies to enable sustainable clothing via durability standards or others.

4) BlockChain technology and AI can be used to scale life cycle management assessments for each product and improve footprint transparency.

5) AI can be used the detect and track green washing claims.

The size of the opportunity is immense, so why haven’t we started using AI to accelerate the circular economy? To finish off, I summarize my own perspective, and that of other industry experts, on the challenges for the adoption of AI in circularity and mitigation.

Challenges and opportunities for AI and the circular economy

As discussed in this article, there are many potential opportunities for AI for fashion to intersect with climate mitigation. However, there are significant challenges to adoption of circular fashion powered by AI.

The first challenge is that the field is underdeveloped both in terms of data availability and AI research. I have identified almost no academic research at the intersection of AI for fashion and climate mitigation. There are currently no open-source datasets that can be used. The extensive use of AI in circular fashion requires good data.

Wide-spread implementation of AI for circular fashion requires gathering large amounts of high quality data. This set of barriers presents an opportunity for academics who are interested in novel research directions and data providers interested in being a trusted source for circular fashion data.

Another challenge is that corporations are often unwilling to stop using trusted, yet inefficient, methods in favor of AI, as in the case of life cycle analysis. Currently, life cycle analysis is a manual, data-intensive process that can be performed for a few hundred products a year. The fashion sector produces over a million new products every season, which is a huge opportunity for the use of AI. However, machine learning-based sustainability claims are considered untrustworthy, and a potential source of greenwashing, preventing them from widespread adoption despite being potentially more accurate than manually input claims.

Corporations must also contend with a changing regulatory environment. European regulatory frameworks will soon require companies to provide detailed reports of the carbon footprints of individual products and may require textile companies to properly treat used textiles. This will put enormous pressure on the data and information systems of fashion companies and provide an opportunity for the implementation of new technologies, like blockchain and AI, to provide scalable and reliable information carbon footprint at the product level.

A final challenge relates to the impact of circular fashion on consumer behavior. Product designs powered by AI could shift consumption towards longer lasting products. However, this may increase consumption, an example of Jevons paradox, in which an action taken to address climate change actually increases emissions. The desire for reduced consumption due to circular fashion is complicated by the need for the fashion companies to be profitable. Further sustainability considerations are required to ensure that as AI makes fashion supply chains leaner and products more affordable and durable, consumers will change their consumption to decrease the climate impact of their fashion.

While these challenges can seem daunting, similar challenges have been overcome in industries such as energy, agriculture, and forestry. In these industries, AI helps reduce overall emissions at scale. The potential benefits of using AI to power circular fashion are enormous and the challenges manageable. The time for circular fashion powered by AI has come.

Getting your academic resume ready for industry

2022-12-02T00:00:00+00:00

Are you a researcher looking to find new opportunities in industry? Have you read guides on how to convert your CV to a resume and showcase your transferable skills, but you’re still not getting there interviews you want?

This guide will show you how to take your resume from “good” to “great.” To do this, we’ll show you how to focus less on the signs of academic success, and more on the needs of industry.

Starting point: the academic resume

I talk to a lot of PhDs, postdocs, and professors who are interested in leaving research for industry (specifically, climate data science and/or management consulting). This means I see a lot of smart, careful people get rejected for jobs they are qualified for because they aren’t attuned to the needs of their potential industry colleagues.

Their resumes often look like this:

Example of a scientific resume that has not yet been tailored for industry positions. Created by Kelly Kochanski.

This example resume is strong, but not good enough

This resume successfully demonstrates that the candidate has many educational, technical, and scholarly qualifications that will be relevant for jobs in their field:

Annotated example of a scientific resume that has not yet been tailored for industry positions. Created by Kelly Kochanski.

However, many elements of the resume are unconvincing:

Annotated example of a scientific resume that has not yet been tailored for industry positions. Created by Kelly Kochanski.

Our goal is to turn this into a resume that shows this candidate is not just qualified, but outstanding.

Tips to fix common resume weaknesses

1. Make your experience outstanding

Every line in your “experience” section should show that you did something better than a mediocre researcher would have done in your situation. If you’ve read resume and interview guides, you’re probably familiar with the situation-action-result format. In research terms, this could look like:

Conducted research on problem X
Applied method Y
Demonstrated result Z

This research experience can be discussed in a mediocre way:

Researched the distribution of mermaid fossils on the seafloor
Developed an extension to a geospatial mapping package, SQUELCH
Demonstrated that current fossil distributions imply a larger population of Paleozoic mermaids than was previously thought

This example could have been written by a student who tagged onto their advisor’s research project, wrote some messy code with no users, and reached a low-impact research conclusion. The same experience can also be discussed in a way that shows the candidate is excellent:

Built Bayesian statistical models of mermaid populations using the AncientSea dataset
Extended a geospatial mapping package, SQUELCH, with Python and ArcGIS. Published my code as part of SQUELCH 3.0 (2000 users)
Demonstrated that Paleozoic mermaid populations were higher than previously thought, leading to two peer-reviewed publications and coverage in the ‘See Sea Science!’ magazine

This example shows that the student has relevant skills and datasets, builds tools that other people actually want to use, and is able to communicate with both researchers and the public.

We’re human, so we’re not outstanding all of the time. Pick 2-5 situations where you did especially good work. Describe those achievements in compelling detail, and cut the rest. Good stories are intriguing. You will be asked questions in interviews, so don’t make things up.

2. Show off your soft skills

Many researchers’ resumes fail to demonstrate soft skills: teamwork, communication, organization, management, mentoring, public speaking, and leadership. These skills are generally valued more highly in industry than in academia, and are critical to almost every job. To show your soft skills, add 2-3 interpersonal stories to your ‘Experience’ section. These should follow the same format as the technical examples above.

Again, there’s a mediocre way to discuss soft skills:

Collaborated with international researchers

This example could be written by someone who tagged along passively on a big collaboration, or even by someone abrasive who made the collaboration worse. The same experience can also be discussed in a way that shows the candidate is excellent:

Realized that the largest failure risk in this research project was poor communication with our international collaborators
Organized short workshops to foster a closer sense of connection within the group
Created a safe, connected working environment where we were able to communicate freely. This led to a smoother work experience, and the group later hosted me for a summer at Cool Foreign University

This example shows that the candidate is thoughtful, aware of social dynamics, good at organizing events, and liked by their collaborators. Your soft skills experience might come from: teaching a challenging course, mentoring students through a challenging moment, organizing a workshop, initiating a new collaboration, engaging non-academic stakeholders, doing outreach, negotiating funding, or leading initiatives in a research organization.

3. Make your awards section sparkle

Academia – especially US academia – gives out a lot of awards.

Most award sections include a couple of gems:

Highly Competitive Fellowship (~$300,000)
Best Speaker Award (awarded to 2 of 50 speakers)
Outreaching/Teaching/Leadership Awards

Mixed in with some rocks of uncertain value:

Tiny University Scholarship ($1000)
University Conference Travel Grant ($500)
Academic Performance Award (x3)

Your “gems” are your most competitive awards, your most financially valuable awards, or awards that demonstrate skills not proven elsewhere in your resume (usually your soft skills). To make the gems stand out, you can put them at the top of the section, use bold font, or call the section ‘Selected Awards’ and drop the “rocks”.

4. Pare it down

Many achievements that are hard-won and dear to researchers do not demonstrate that you can do non-academic work. These should be removed to keep readers focused on your most relevant experience. This will almost certainly require you to remove several sections that are academically prestigious:

Don’t list publications, and certainly don’t give full citations. Publication quality is hard for people outside of your field to judge.
- Do use lines like, “Demonstrated result Z, leading to two publications in peer-reviewed journals” in your experience section.
- Do include evidence of impact, such as references in mainstream media, public policy, or improvements in the work of other researchers.
Don’t list conference talks. Talk quality is hard for anyone who wasn’t there to judge.
- Do give evidence that you’re a skilled public speaker, such as “best speaker” awards, or evidence of positive receptions to talks.
- Do use high-profile talks as evidence of success, e.g. “I was invited to present this project as spotlight talk at the 200-person Iapetus Ocean workshop.”
Don’t list membership of societies or committees. Many committees achieve little.
- Do use committee service as evidence of leadership skills, e.g. “Introduced a safer and more private procedure for addressing student complaints, receiving thanks from the undergraduate diversity committee.”
Don’t list courses taught. Teaching skill is hard to judge from course titles.
- Do use teaching as evidence of hard skills, e.g. “Re-wrote lectures for Computational Ocean Science class to include a Python programming module, improving student evaluation scores from 4.5 to 6.0/7.”
- Do use teaching as evidence of soft skill, e.g. “Reached the highest enrollment demand for any elective in the department” or “Three students I mentored earned competitive awards for their research”
Don’t duplicate material. If you have 5 lines discussing your experience with Ocean Science or Python, remove the weakest to make space for other material
Don’t use a summary or objective statement if you can avoid one. Opinions differ on this; I believe that most summaries duplicate material given elsewhere

Finally, read your resume defensively. Are there any lines in there that could have been written by a mediocre researcher? Take them out.

Summary

To improve your resume, make sure that every line shows you are outstanding. This means:

Each item in “experience” shows you did something interesting that a mediocre researcher would not have done
2-3 items in “experience” show you use soft skills that an abrasive researcher would not have
Each of your awards adds something substantial
Each point that does not demonstrate relevant skills is removed

Endpoint: a better resume

Here’s my (quick and imperfect) improved version of the resume we started with:

Example of an academic resume that has been well-tailored for industry positions. Created by Kelly Kochanski.

Think you can do better? Edit the example here and send your version to kelly (dot) kochanski (at) gmail (dot) com.

Using Machine Learning to Track International Climate Finance

2022-11-09T00:00:00+00:00

International climate finance is a key ingredient for global action to curb global temperature rises and prepare societies for the already irreversible effects of climate change. At the 2009 Copenhagen summit, high-income countries committed to mobilize US$ 100 billion annually from 2020 onwards to support mitigation and adaptation in developing countries. This promise was a crucial step to get lower and middle income countries to define national emissions reduction targets and led to the Paris Agreement where, for the first time, all countries committed to climate action.

The latest report by the OECD shows that, in 2020, contributor countries failed to meet the US$ 100 billion climate finance target. This weakens trust of recipient countries and complicates the current climate negotiations in Egypt (COP27), where countries are starting to define a new climate finance target for the period after 2025—the “new collective quantified goal.”

Tracking global flows of climate finance is difficult. Countries and institutions self-report the projects and corresponding finances which contribute to the US$ 100 billion target. However, these contributors have different methods for assessing what counts as climate finance, and the reporting process is not transparent. Moreover, researchers and NGOs have found that many projects are misreported and should not be counted as climate finance.

Analyzing project descriptions with natural language processing

In our new study, Florian Egli, Anna Stünzi, and I developed a natural language processing model (ClimateFinanceBERT) that identifies climate finance projects based on their textual descriptions and classifies them into granular climate finance categories (e.g., solar energy or energy efficiency). Using ClimateFinanceBERT, we analyzed 2.7 million descriptions of bilateral development projects from 2000 to 2019. The model classified 80,023 of these projects as climate finance (52% adaptation and 48% mitigation projects), totalling US$ 80 billion.

For the period after the Paris Agreement (2016–2019), our estimates are about 64% lower than the officially reported project assessments. These findings show a great disparity between promised and delivered climate finance and support claims that contributor-reported numbers may be inflated.

The study also provides important take-aways on how machine learning could help address major challenges in climate finance accounting. The machine learning based approach offers three benefits: flexibility regarding the scope of climate finance, consistency in the evaluations of textual descriptions, and scalability that enables cost- and time-effective analysis regardless of the number of documents.

Flexible scope: ClimateFinanceBERT allows for a flexible definition of climate finance, because its assessment is based on a set of granular categories, like solar energy or energy efficiency. In the paper, we focused on a strict conception of climate finance where we only counted projects with a primary focus on adaptation or mitigation as climate finance. However, our model also categorizes other types of projects with co-benefits for the climate, such as biodiversity or sustainable land use projects. This “bottom-up” approach—using a collection of granular categories instead of a single yes/no classification—offers the flexibility to extend the scope of climate finance by including more categories with co-benefits for the climate.
Consistent evaluation: ClimateFinanceBERT provides a consistent evaluation of climate finance that can easily be replicated. The current reporting of climate finance lacks such consistency, on the one hand because reporting organizations differ in how they assess climate relevance, and on the other hand because human annotators vary considerably in interpretation and individual judgment. The lack of consistent reporting makes it difficult to compare numbers between different contributors and to provide a transparent and comprehensive picture on global flows in international climate finance.
Scalable classification: Replacing human evaluations with machine learning classification could open a major bottleneck in climate finance accounting: time and cost. Classifying individual projects via human annotators takes considerable time, which means that evaluation is done only once and, most likely, by only one person. Changing decision paradigms ex-post or double checking individual evaluations is simply too time consuming and expensive. In contrast, machine learning algorithms are scalable. Depending on the hardware, it takes between several minutes and a few hours to classify millions of textual project descriptions with ClimateFinanceBERT.

What role can machine learning play in climate finance accounting?

Despite its benefits, a machine learning-based approach cannot substitute entirely for human evaluations.

Climate finance is complex, and so is the task of classifying it. Projects related to climate change adaptation are highly context-specific. They may require expert knowledge of other external factors, which are not provided in the project descriptions, to classify them correctly. For example, water and sanitation projects are more relevant for climate adaptation in regions affected by droughts.

Furthermore, the complexity of climate finance means machine learning classifiers may misclassify some infrequent and underrepresented types of projects. This risk is intrinsic to supervised classification: a classifier trained on one set of project descriptions must generalize to new descriptions it hasn’t seen during training.

In our study, we trained our classifier on 1,500 climate finance projects and then classified 2.7 million project descriptions. To mitigate blind spots, we:

used a stratified approach to create our training set by sampling project descriptions from different sectors (e.g., energy, agriculture) and with different climate finance tags by the contributors (e.g., mitigation, adaptation);
used a model that was already pre-trained on three million climate related texts before training it on project descriptions from international climate finance; and
conducted extensive validation checks, testing our trained model against existing manually annotated data sets from other climate finance studies and against expert judgment in user studies.

While machine learning will not be able to replace human evaluators, it is a helpful tool that can complement human assessment by double-checking reported projects, simulating outcomes for different climate finance scopes, and analyzing how climate finance is distributed over different sub-categories.

Double-checking reported projects: Machine learning models could double-check reported projects and highlight potential misclassifications or relevant projects that have been overlooked by reporting contributors. In our study, we found both substantial over- and underreporting of climate finance projects. Alternative suggestions by the machine learning tool could be manually checked through an expert verification process. Contributors could use the tools to cross-check their own efforts. Overall, algorithmic double-checking could improve trust in the current reporting at minimal cost and effort.
Simulating outcomes for different climate finance scopes: Machine learning models could be used to analyze how different scopes of climate finance accounting affect overall financial contributions and relate to climate finance targets. This would create a more nuanced picture of international climate finance, informing negotiations and nurturing climate ambitions.
Analyzing climate finance distribution: The bottom-up approach of the machine learning model could help analyze how climate finance is distributed over different sub-categories, such as extreme weather, renewables, or energy efficiency. This would allow contributors, recipient countries, and other stakeholders to identify categories that are under- or overrepresented in international climate finance, helping them coordinate between projects and effectively allocate financial resources.

While machine learning is not a silver bullet, it allows parties such as contributors, recipients, and NGOs to review climate finance contributions based on consistent criteria and a flexible scope. This would enable all parties to discuss targets for climate finance and climate action at eye level. Ultimately, tracking international climate finance with machine learning could help restore trust in the global community, which is essential for agreeing on credible climate finance targets and translating financial support into effective climate action.

Bridging the Gap Between Academia and Industry In The Energy Sector

2022-10-24T00:00:00+00:00

Despite the wealth of cutting-edge research produced in academia, it is difficult to extract these learnings and embed them into the operational, day-to-day activities of industry. This innovation chasm between academia and industry is known as the “Valley of Death,” where many good ideas are lost.

Data science innovation within the energy sector loses many great ideas in the valley of death. In this post, we provide insights from a recent report which investigates the challenges and opportunities in sharing knowledge and skills from academia to industry.

The challenges and opportunities of the valley of death were highlighted to Energy Systems Catapult, a UK not-for-profit technology and innovation centre set up to support innovation in the energy sector, after a meeting with the innovative flexibility providers Arenko Ltd. Arenko, an operator of flexible assets in the UK energy market that focuses on algorithms and automation, had struggled with accessing and reproducing the valuable research contained within academic journals, despite their need to keep up with the latest state-of-the-art models.

Identifying the Valley of Death

To understand the landscape of academic impact in the energy industry, especially within the fast-moving data science domain, we conducted desktop research and interviewed over 30 professionals from both academia and industry. These professionals provided insight into their difficulties in collaborating and transferring knowledge. They also shared insights about what worked well, or poorly, in their interactions across the valley of death.

A common theme in our interviews was disappointment with collaborations between academia and industry. A driving factor in this disappointment is that the two sectors’ timelines and objectives are misaligned. Industry expects relatively short turnover (months), whereas academic projects are usually much longer (years). Additionally, their success measures and criteria are often very different. For example, in academia a common objective is a large number of citations on publications in high-impact journals, whereas industry focuses on other KPIs such as financial targets or levels of engagement.

When developing projects together, academia and industry often don’t fully understand the differences in work culture and end up choosing suboptimal ways of working together. As a result, objectives and outputs deviate from the original proposal, with academics often focusing on more theoretical outputs, leading to less application to the original business case. This creates many missed opportunities, not just for innovation, but for identifying and recruiting the best academic talent.

To supplement the interviews, we also reviewed the literature on a specific area of data science research in energy systems: electricity price forecasting. Arenko’s domain expertise allowed us to understand what parts of these academic publications could be used in industry settings. The literature review showed a disconnect between industry needs and academic practices. The methodologies outlined in most papers are often not clear enough to enable reproduction, and without open data and/or code, it is often virtually impossible to verify the results and methods. For industry practitioners with limited time and resources, these instant roadblocks make it far less likely that a company will apply academic research.

The interviews and the literature review showed an important training gap in coding, a skill set that is desperately needed given the rapid digitalisation of the energy sector. Although data science courses develop rigorous skills in modeling and advanced algorithms there is very little content on applying the latest programming approaches, e.g., code review, structure, or version control. Of ten data science-focused master’s programs from major universities in the UK, we found only two that had dedicated introductory courses to programming and none of them had any dedicated intermediate or advanced courses. In other words, unless self-taught, such graduates entering the workforce will be lacking in the skills to develop operational code in industry teams. Discussions with those in the energy sector have highlighted this as a major difficulty when recruiting new data science talent.

Making Impact

These are just some of the issues we discovered through our research. To better address the long-standing misalignment between academia and industry our reports collate many tools and approaches to support more fruitful and fulfilling academic impact in industry.

One of the most effective, but perhaps difficult, ways to improve collaboration is better understanding of each other’s cultures, businesses, and objectives. Is it more effective to collaborate through master’s projects, PhDs, or postdocs? Can the project be designed to achieve satisfactory outcomes for both sides: publications for academics; useful tools and insight for the company?

One approach to improved collaboration is to move away from short-term projects and develop longer-term strategic relationships between business leaders and academics. This can give clearer objectives for collaboration and support postdoctoral researchers, a valuable and often overlooked academic partner. All too often, postdocs, who are often the most versatile and effective researchers, are given short-term contracts with extremely large workloads. Longer-term contracts will not only attract the best academics but also help maintain continuity and quality in the relationship.

Another recommendation is for academic research to focus on open and reproducible research. All models should be trained on open data sets and verified using common, simple benchmarks. Code should be treated as a valuable research output to incentivise code sharing within academia. This can improve coding skills, serve as an advertisement to prospective employers, and help support reproducible research and innovation. Platforms like Hugging Face which share trained code and data are becoming more popular and should be adopted by the energy data science community.

Data Science Competitions co-organised with industry have proven to be incredibly useful platforms for supporting open science and the development of common benchmarks to give innovators a head start. They can also help identify some of the most powerful methods and input features, helping crowdsource solutions for some of the biggest challenges facing the energy sector. By having participants share repositories, these competitions also encourage the sharing of new and novel data sets.

Finally, applied academic research depends on the real-world challenges and problems facing the industry. Industry plays a key supporting role for new innovative research. Companies must share their experience, challenges, and data if they want research to solve their relevant issues and make a difference in the real world. Industry needs to engage proactively with universities to make sure they are solving the actual problems they are facing and can utilise their domain knowledge and experience.

Similarly, academia must better utilise and listen to industrial experience and knowledge in shaping their courses. Industry can play an important role in developing the coding skills of the future graduates. Software development moves at a faster pace outside of academia, so their input is vital to ensure training that is fit for purpose!

Finding More Information

Our Report, Data Science: From Academia to Industy

The learnings and challenges we outlined above and more can be found in the series of reports we developed. The main report considers the core aspects and challenges of bringing innovative ideas from academia to industry in data science. Four supplementary reports describe mechanisms which can support this impact: Collaboration, Accessible and Reproducible Research, Code Development for Academics, and Industrial Support for Academia.

We also invited a number of experts in these topics to share their views and insights in a recorded launch event.

We do hope you enjoy these resources and if you are interested in further discussions, or have ideas to extend or enhance this work, we’d love to hear from you! Contact us at Energy Systems Catapult: https://es.catapult.org.uk/contact/

The Green Last Mile Project: Accelerating Cargo-Bike Logistics in Cities

2022-10-11T00:00:00+00:00

Light goods vehicles (LGV) are one of the leading polluters in cities due to their extensive use in the last mile of delivery. Recently, cargo bike logistics has been put forward as a competitive, zero-emission candidate for replacing LGVs in urban areas, with experts usually estimating that cargo bikes can replace 25-50% of van deliveries. Today the percentage of last mile deliveries made by cargo-bike remains anecdotal, mainly pioneered by a number of smaller logistics operators (our partners Pedal Me and Urbike being prime examples of that), but is gaining increasing attention from the logistics industry.

Cargo-bikes are a multi-solution on the path to sustainable and humane cities. They cut pollution, decongest cities, emit a tenth of the CO2 emitted by an electric van, and take much less space on the roads, all this while moving significantly faster in dense urban areas. Due to their faster speeds, shorter parking times, and more efficient routes across cities, they can out-compete traditional van logistics when operated effectively.

Despite these competitive advantages in dense urban areas, their widespread adoption by logistics operators is limited. This is primarily due to the lack of inexpensive ways of accurately evaluating their impact on the cost of business. By modelling the relative delivery performance of different vehicle types across urban micro-regions, machine learning can help operators evaluate the business and environmental impact of adding cargo-bikes to their fleets.

Our project is developing accurate models of navigation time and service time (e.g. cruising for parking, unloading, walking) for different vehicle types during urban delivery runs. Using Uber’s H3 index to divide the cities into hexagonal cells, and using OpenStreetMap tag data, we are studying how urban context affects delivery performance for different vehicles. Our models will enable key stakeholders to run feasibility studies for optimising and diversifying their fleet composition in a cost-effective manner.

Cycle-logistics as a sustainable solution to urban transport has not been a focus of the machine learning community, where solutions such as autonomous vehicles, delivery drones, and robots get attention and funding, despite their uncertain feasibility and impact.

A significant uptick in cycle logistics would also make transport cycling much safer and thus encourage people to move away from cars through 1) removing vans from the roads, 2) providing commercial incentives for cycling infrastructure and 3) fostering a phenomenon known as ”safety in numbers” (the more people cycle, the safer it is, thus the more people cycle). In cities, 50% of car trips are shorter than a 20min cycle. Shifting away from short car trips could have wider repercussions on transport GHG emissions.

[Pedal Me](https://pedalme.co.uk/), a cargo-bike logistics operator in London, UK

About the Partners

The Green Last Mile Research Group

Our research team is composed of Navish Kumar (IIT, Kharagpur 🇮🇳), Max Schrader (University of Alabama 🇺🇸), Maria Astefanoaei (Dasya Lab, ITU, Copenhagen 🇩🇰), Akash Srivastava (MIT-IBM AI Research Lab, Cambridge, MA, 🇺🇸), Kai Xu (Edinburgh University / Amazon Research, NYC 🇺🇸), Esben Sørig, Soonmyeong Yoon and Nicolas Collignon (Kale Collective, London 🇬🇧). The research is being conducted in partnership with two expert cargo-bike logistics operators, Pedal Me in London 🇬🇧 and Urbike in Brussels 🇧🇪, as well as Larry vs Harry, a leading cargo-bike manufacturer in Copenhagen 🇩🇰.

The Team

From left to right and from top to bottom, Soonmyeong Yoon, Esben Sørig, Nicolas Collignon, Kai Xu, Max Schrader, Navish Kumar, Akash Srivastava, and Maria Astefanoaei

Estimating the Ice Volume of All Glaciers in High Mountain Asia With Deep Learning

2022-10-03T00:00:00+00:00

Estimating the ice volume of the Earth’s glaciers is a grand challenge of Earth system science. Besides being a critical parameter to model glacier evolution, knowledge of glacier volumes is fundamental to forecasting global sea level rise as well as available freshwater resources.

Glaciers are very sensitive to climate change. Under atmospheric warming induced by human forcing, with few exceptions glaciers have been retreating worldwide at unprecedented rates. Glacier melting is estimated to be responsible for 22% of the sea level rise observed during 1971-2018, with far reaching implications for the human population living in coastal areas.

Ice mass loss from glacier shrinkage has impacts on water availability for an estimated global population of 1.9 billion people living in or depending on glacier sourced freshwater .

In the Himalayan region in particular, melting glaciers poses a threat to water security for more than 1.4 billion people who depend on glacial basins for water , in addition to exacerbating the risk of floods that are wrecking the region, driven by increasingly more powerful monsoon rainfall. Overall, the future pathways for freshwater availability are uncertain and heavily depend on the upcoming climate trajectories, making improved glacier ice volume estimates in this region an urgent priority.

In a recent study, a group of researchers reported that the total glacial ice volume in High Mountain Asia is 37% higher than the previous consensus value, highlighting the challenges of traditional glacier modeling approaches and the need for more accurate and less uncertain ice volume reconstructions. With ICENET we explore the potential of the immense pool of satellite images and recent advances in deep neural networks to advance the knowledge of glacier ice volume in High Mountain Asia.

Top: digital elevation map of a 1° x 1° region over the European Alps. Bottom: the same region sampled to create the training dataset needed for the deep learning model.Photo credit:Niccolo Maffezzoli.

Remote sensors mounted on satellites have been silently monitoring the extent and changes of the glaciated surfaces of the Earth for decades. The accumulated vast amount of data produced by these instruments are difficult to exploit by traditional data-analysis pipelines, and at the same time the perfect playground for neural networks. Among the products delivered by satellite sensors is data of the Earth’s surface elevation, including glaciers. To calculate their ice volume, however, the missing piece of information is the bedrock topography, i.e. the spatial distribution of the elevation of the Earth’s crust beneath glaciated regions. Our model will attempt to infer such glacier bedrock topographies. As an elevation satellite product, we use 1-dimensional digital elevation maps of the Earth surface.

We approach the problem leveraging generative deep inpainting architectures trained on extensive portions of de-glaciated mountain regions. Inpainting networks are used to fill missing patches within an image. In our case the training images are such elevation maps of mountain regions, while the missing patches, i.e. the targets of the network, are represented by the glaciated areas. The network will thus be trained to reconstruct glacier underlying topographies, and in turn their volume by the difference with the surface elevation.

With the generative models developed over the course of the ICENET project, we hope to reduce the current uncertainties of ice volumes in High Mountain Asia, and at the same time advance knowledge of glacier modeling with the help of AI techniques.

The Team

From left to right: Niccolo Maffezzoli (National Research Council, Italy),Eric Rignot (University of California Irvine) and Carlo Barbante (National Research Council, Italy)

The Institute of Polar Sciences (CNR-ISP, Venice, Italy) is part of the Italian National Research Council. The mission of the ISP is to improve our understanding of the climatic changes taking place in the Arctic and Antarctic environments and possible future developments at both polar and global levels. Our studies address research issues related to both the chemical/geochemical and physical aspects of the poles using a multidisciplinary approach to protect these vulnerable extreme environments. ISP also has a long standing record of research on mid-latitude glaciers, especially in the European alpine region.

The Department of Earth System Science (ESS) at the University of California, Irvine focuses on how the atmosphere, land, and oceans interact as a system, and how the Earth will change over a human lifetime. At ESS, researchers combine observations from remote sensing platforms and field data with numerical modeling to understand the physical processes controlling the response of the ice sheets to climate change, and to reduce the uncertainties of projections of the future contributions of the Greenland and Antarctic Ice Sheets to sea level rise regionally and globally over the coming centuries.

Tackling Climate Change with Machine Learning Workshop at NeurIPS 2022

2022-09-12T00:00:00+00:00

We are thrilled to bring another edition of Climate Change AI’s “Tackling Climate Change with Machine Learning” workshop at the upcoming NeurIPS 2022 conference! This workshop series is part of events CCAI has organized at major machine learning conferences such as ICML (2019, 2021), ICLR (2020) and NeurIPS (2019, 2020, 2021).

For this iteration of the workshop, the keynote talks and panel discussions features leaders from industry, academia, and government who will be particularly focused on exploring the theme of climate change-informed metrics for AI, focusing both on (a) the domain-specific metrics by which AI systems should be evaluated when used as a tool for climate action, and (b) the climate change-related implications of using AI more broadly.

We are currently accepting short papers (up to four pages) and proposals (up to three pages) at the intersection of climate change and machine learning until September 20th, 23:59 AOE.

We are also excited to announce a special issue for ‘Tackling Climate Change with Machine Learning’ where for the first time in this workshop series, interested authors will be invited to submit a full journal length article for publication in the Environmental Data Science journal.

Important resources regarding the workshop and special issue:

This post contains the high-level requirements for successful submissions to complement the informational webinar. We encourage you to explore previously accepted papers search tool on our website. The contributions featured at our workshops are non-archival in nature, and the authors are free to subsequently publish or present their work elsewhere, including at the special issue.

We have summarized below some key expectations from papers and proposals.

Successful papers (up to four pages, excluding references and appendix):

Contain research with results
Can be a work-in-progress
Need not contain complex ML if the application is compelling

We encourage submissions that include dataset publication and describe deployed technologies.

Proposals should include (up to three pages, excluding references and appendix):

A detailed plan for future work
An explanation why solving your problem matters, how you plan to solve it
Identify relevant stakeholders that play an important role in your proposed solution

Additionally, successful papers in our past workshops clearly explain the motivation of their study/analysis and the impact of the work therein. These submissions avoid jargon and explain/justify the use of ML. Lastly, the successful submissions are expected to:

Illustrate the link: Explain the way in which your results could help address climate change. Theoretical work with a multi-step link is fine, as long as it is clearly communicated.
Consider relevant stakeholders: Whose actions are you ultimately trying to influence? What insights would they find valuable?
Outline key metrics: Quantitatively or qualitatively explain why solving your problem is important and how well you are doing versus other methods.
Be clear and concise
Convey the big picture: Explain how your work aligns with CCAI’s mission to “empower work that meaningfully addresses the climate crisis”

We encourage you to explore our website for more information at https://www.climatechange.ai/events/neurips2022 and reach out to us at climatechangeai.neurips2022@gmail.com with any questions. We look forward to your submissions and will hopefully see you (virtually) at NeurIPS!

Members of the organizing team,

Peetak Mitra (Xerox PARC), Maria João Sousa (IST, ULisboa), Mark Roth (Climate, LLC), Ján Drgoňa (PNNL), Emma Strubell (Carnegie Mellon University), Yoshua Bengio (Mila, UdeM)

Copy of the Tackling Climate Change with Machine Learning 2022 Poster

Detecting Flooding in Fiji’s Croplands

2022-09-06T00:00:00+00:00

In Fiji, flooding from tropical cyclones and storms frequently causes considerable damage to key croplands, leaving impacted farmers in need of assistance. For example, following tropical cyclone Cody in January of 2022, an initial assessment estimated damages to the agricultural sector at over US$4 million¹, and more than US$1.3 million in relief was ultimately paid in assistance to farmers. Flood risk is only expected to increase with climate change: recent studies project millions more dollars in agricultural losses and damages. As such, responding to floods has become a key component of Fiji’s climate change policy.

To serve that need, our project will develop a new algorithm to quickly generate cropland flood maps following tropical cyclone events in Fiji. Our tool will use machine learning to classify flooded croplands using satellite images. We will also develop a web application to deliver these flood maps to agricultural recovery and relief stakeholders. Key features will include predicting flood occurrence in areas initially obscured by cloud cover, detecting flooding in small-scale and remote croplands, and updating flood extent predictions using temporal profiles of crop growth. Operationally, after a tropical cyclone, these cropland flood maps will be accessible through a mobile geospatial data collection app, QField, to help Fiji’s Ministry of Agriculture staff conduct damage assessments.

An example classified cropland flood map viewed as a layer in QField mobile GIS. The map can assist with damage assessment surveys and images after storm flooding in Fiji. Basemaps: OSM Bright by OpenMapTiles and ESRI World Imagery.

Using this approach to classify images, an archive of cropland flood maps will be generated from historical satellite images and made available for practitioners and researchers. This archive will be updated as new tropical cyclone events occur. It will contribute to a growing understanding of patterns of flood risk and impacts across Fiji’s diverse cropping systems.

Our project builds on previous work by our team and others. In a previous project, our team collaborated with stakeholders in Fiji and Tonga to develop workflows and apps for field-based landscape mapping. We also developed classification methods for estimating inter-annual land cover and generating cropland maps for Fiji from stacks of satellite images. Other recent progress includes workshops with team members and officials from Fiji’s Ministry of Agriculture to identify how flood maps can best augment the Ministry’s damage assessment activities.

Classroom and field-based training sessions have also been run to build damage assessors’ capacity to use QField for in-field data collection. QField’s functionality enables flood maps to be viewed on a mobile map display, which is integrated with digital forms for damage assessment surveys, and can guide damage assessors to sites of flooding. Fieldwork has also been conducted in the Rewa Delta, with staff from The University of the South Pacific and the Ministry of Agriculture testing data collection apps for flood damage assessment and collecting ground truth data.

Over the coming months, we will work towards finalising the machine learning workflow that generates flood maps to assist with the critical disaster response task of assessing flood impacts on Fiji’s croplands.

About the partners

The UWA School of Agriculture and Environment (SAGE) and The Centre for Water and Spatial Sciences (CWSS) are based at University of Western Australia. SAGE has expertise in natural resource management and agricultural and environmental sciences. CWSS is a cross-university centre hosting multidisciplinary researchers and industry collaborators with expertise in geospatial and water sciences.

The University of the South Pacific is one of the Pacific region’s largest regional universities, home to 14 campuses spread across its 12 member countries. The USP’s School of Agriculture, Geography, Environment, Ocean and Natural Sciences (SAGEONS) is critical to the development and research within the region. It is actively involved with climate change research in agriculture, especially given the impacts of climate change-induced floods and natural disasters that affect Fiji and the Pacific.

The Ministry of Agriculture (Fiji) is responsible for enhancing Fiji’s food production and income security through agricultural sector growth. The ministry works to ensure the protection of crops and croplands in Fiji. The Land Use Section within the ministry’s Research Division works with partners and stakeholders to develop broad and effective solutions to climate change and the impacts of flooding on the agricultural sector.

The School of Geosciences at The University of Sydney is a dynamic group of interdisciplinary researchers with expertise in geocoastal systems, geography, geology, and geophysics. The school tackles key issues facing society, including climate change, resource management, and sustainability.

Team: Solomoni Nagaunavou, Nathan Wales, Renata Varea, John Duncan, Eleanor Bruce, Kevin Davies, Bryan Boruff, and Diana Ralulu.

Notes

Tropical Cyclone Cody Initial Damage Assessment Report, Ministry of Agriculture. ↩

Scaling Climate Smart Shrimp in Southeast Asia using GIS and Computer Vision

2022-06-16T00:00:00+00:00

Mangroves are critical for coastal adaptation to climate change, providing protection from rising sea levels and storm surges while also sequestering carbon. However, like many other forests around the globe, mangroves are threatened by deforestation. One of the main drivers of mangrove deforestation is shrimp farming, especially in Southeast Asia, where it has driven 30% of forest loss and land use change in the region.

To support shrimp production while restoring mangroves in critical areas, Conservation International (CI) has initiated the “Climate Smart Shrimp” (CSS) program in Indonesia and the Philippines. In Indonesia, CI is working with its in-country partner Konservasi Indonesia (KI) to pilot the CSS program this year. The program will apply a novel “green-gray” engineering approach that can restore mangroves while simultaneously helping farmers intensify their production. CI estimates that this approach could restore up to 1.7 million hectares of mangrove forests worldwide, while producing just as much or even more shrimp.

The “Climate Smart Shrimp” Program combines responsible intensification of shrimp farming with mangrove restoration.

Codifying Domain Expertise with Multi-Criteria Decision Analysis

The key starting point for the CSS program is identifying the most suitable shrimp and other aquaculture farms at the largest scale possible. However, scoping these farms on the ground for one country alone can take months or even years, depending on the resources available.

Using a Multi-Criteria Decision Analysis (MCDA) approach, the consortium developed a set of filtering criteria supported by open geospatial data and satellite imagery to rapidly identify the most suitable aquaculture sites for further on-site validation. These criteria include characteristics like proximity to roads and populated areas, proximity to historical and present mangroves, pond size, and slope and elevation. Because the criteria are easily interpretable, they encourage feedback loops and quick iteration to continuously improve the filtered results.

By the end of the CSS program, the goal is to find the top 40,000 most suitable hectares in Southeast Asia. The results will be used by CI and their partners (including KI) to engage local communities and policymakers, as well as to raise funds from private investors to further strengthen and scale out the CSS program. Ultimately, the program supports the Global Mangrove Alliance’s 2030 goal of increasing mangrove cover globally by 20% while addressing challenges in food and job security and protecting coastal communities in disaster-prone areas from the impacts of climate change.

The CSS program is set to take place this 2022 and will produce open-source code, datasets, interactive maps, and research products. Stay tuned for more updates or reach out to Conservation International (community@conservation.org), Koservasi Indonesia (asupriatna@konservasi-id.org), or Thinking Machines (data-for-good@thinkingmachin.es) for more information.

About the Participants

Arizona State University (ASU) is a public research university in Phoenix, Arizona founded in 1885. ASU houses the Julie Ann Wrigley Global Institute of Sustainability and is home to the United States’ first School of Sustainability. ASU’s sustainability work centers on education, research, business practices, global partnerships, and interdisciplinary solution initiatives.

Conservation International is a global non-profit organization established in 1987 that works to protect nature through fieldwork, working with Indigenous communities, governments, and corporations, and innovations in science, policy, and finance. CI’s priorities include climate stabilization through nature-based solutions, protecting the ocean, and expanding planet-positive economies. CI has protected over 6 million square kilometers of ecosystems around the world and currently has offices in over 24 countries.

Konservasi Indonesia is a national foundation that aims to support the sustainable development and protection of critical ecosystems in Indonesia. As the main partner of Conservation International in Indonesia, KI works in partnership with government and other stakeholders to design and deliver innovative nature-based solutions for climate change, using a sustainable landscapes-seascapes approach to create lasting impacts for people and nature.

Thinking Machines Data Science (TM) is a technology consultancy that builds cloud artificial intelligence (AI) and data platforms to solve high-impact problems for large organizations across Southeast Asia. TM’s Sustainability Team is constantly looking for new opportunities where data science can support and strengthen solutions for climate action and social impact. Having worked with a diversity of partners ranging from nonprofits and government agencies to top corporations, the company is also a Google Cloud Platform Partner, Waze Research Partner, and portfolio company of the UNICEF Innovation Fund. TM was founded in 2015 and now has operations in Manila, Singapore, and Bangkok.

Team Members

Project Team members from Conservation International and Konservasi Indonesia (left-right, top-bottom): Ana Plana Casado, Garrett Goto, Hanggar Prasetio, Satya Reza Faturakhmat, Dane Klinger, Ateng Supriatna, and Juliana Tomasouw.

Project Team members from Thinking Machines Data Science (left–right, top-bottom): Anica Araneta, Oshean Lee Garonita, Joshua Cortez, Ren Avell Flores, JC Nacpil, Pia Faustino, JT Miclat.

If you’re curious to see how this project turned out, the CSS team wrote an updated blog post with project outcomes.

Carbon Capture, Extreme Weather, and Disaster Management

2022-05-31T00:00:00+00:00

This past February, CCAI kicked off our first batch of reading groups. Now, a month after the groups wrapped up, I wanted to share what the groups learned and how CCAI plans to iterate on the reading group concept.

As CCAI’s Community Events lead, I was excited to organize the reading groups because I love to learn new things alongside other people. That love stems partly from the social aspect, which CCAI’s online learning resources, content-rich as they are, can’t provide. But I also appreciate the shared accountability of discussing new knowledge with group members—it pushes me to engage a bit deeper with the material than I would on my own.

Our reading groups were led by experts in their respective fields of research. With minimal support from me, these leaders picked the content, structure, and schedule. The goal was for the groups to capture many of the enjoyable aspects about school—interesting topics, open discussions, passionate classmates—while leaving behind the stresses of grades.

Open to everyone and anyone, the reading groups received an abundance of interest. We were excited to offer groups on 3 different topics: Earth Observation and AI for Disaster Management and Relief, Deep Learning and Extreme Weather Events, and Carbon Capture and Machine Learning. Every other week, the members would meet virtually for about an hour to discuss an agreed-upon paper. These meetings were sometimes centered on a presentation, but other times it was easier to discuss the content in an open conversation or by working together through the technicalities of a paper section by section.

I want to thank our three reading group leaders, without whom these reading groups and the ensuing knowledge sharing and growth would not have been possible.

To offer a window into this round of reading groups, three participants and group leaders took some time to reflect on their experiences.

Eli Sherman, who participated in the Deep Learning and Extreme Weather Events reading group, is a PhD student in the Johns Hopkins Computer Science Department studying observational causal inference and machine learning methods for resource allocation. He wrote:

Participating in the CCAI Reading Group was a great experience for me. I’ve spent many years working in the healthcare space and had been trying to find a way to break into the climate change problem space, and this group served as an excellent opportunity to do so. The group included a diverse array of experts, from AI/ML practitioners to those with environmental science expertise but little AI experience. Each week gave me the chance to read papers that helped illustrate that work on climate change is not too far off from my past experience, so my goal of contributing to climate change solutions is, perhaps, not too far out of reach.

Olga Kaiser led the Deep Learning and Extreme Weather Events group. Olga is a senior researcher at NNAISENSE applying deep learning to industrial problems. With a background in mathematics and physics, Olga is interested in improving climate and weather prediction with machine learning techniques, with a particular focus on how causality is integral to understanding complex systems. Olga wrote about her experience leading her group:

I was delighted to join the CCAI by leading the Deep Learning (DL) and Extreme Weather Events Reading Group. We started with the basic statistical concepts of extreme events, causality, and DL. After that, we discussed papers that focus on why causality matters, what challenges we need to address, how precursors to weather events can be identified, and how DL as the core can help and account for causality. During preparations, I had an excellent opportunity to structure the state-of-art literature and learn a lot by finally taking the time to read these papers. It was great to meet and learn from a community that loves sharing and taking in knowledge.

Zikri Bayraktar is a senior research scientist at the Schlumberger-Doll Research Center studying machine learning applications for engineering problems in the field of oil and gas exploration. Zikri spoke to the diverse range of individuals that his reading group attracted:

20 individuals signed up for the Carbon Capture & Machine Learning reading group, calling into our Zoom meetings from parts of the world spanning India, Portugal, Belgium, and the US. We also connected over CCAI’s Circle platform for offline discussions of the selected papers. The topic of carbon capture is young but vast, ranging from capture at point sources like power plants or hard-to-abate industrial facilities to CO2 removal via natural processes in forests, wetlands, or oceans. The open-literature papers we picked for our discussions demonstrated to us how many ways ML can help address challenges in carbon capture, including screening potential solvents for CO2 capture at point sources and classifying or segmenting forests for better quantification of CO2 capture in oceans or wetlands. Our reading group highlighted the importance of working with domain experts on carbon capture, as ML applied to highly specialized domains requires highly accurate methods. Our readings also showed us that there is room for all of us, with our unique expertise and perspectives, to help tackle the challenge of carbon capture with ML.

The wide range of timezones that Zikri mentions was something we had hoped for and celebrated. To know that these groups attracted participants beyond the local communities of our reading group leaders was awesome—but not without obstacles. This diversity came with serious logistical challenges that bottlenecked some of our participation and community building.

We received lots of feedback from both members and group leaders, which we are learning from to create a more engaging experience for everyone interested in future reading groups—even those who cannot make it to the meetings.

On a related note, I am happy to announce some exciting news: our interest form for our next batch of reading groups is now live! We are again offering three groups which will be run similarly to our first batch. One of them, though, is unlike any reading group I’ve been a part of—its discussions will be based on a podcast! 🙂

You can read more about the next round of reading groups and sign up using our interest form. You can also read more by joining the groups’ online communities.

I hope to see you online and/or in the reading group sessions!

Announcing the Climate Change AI Innovation Grants 2022 winners

2022-04-13T00:00:00+00:00

We at Climate Change AI are excited to announce the winners of our 2022 Innovation Grants program. These winners were selected through rigorous peer review by a range of experts in AI and climate change, and will receive USD 1.8M total split across thirteen project proposals.

The projects were chosen from almost 200 submissions from around the world, each proposing impactful, AI-driven approaches for addressing climate change mitigation and adaptation. The awarded proposals span a broad range of application areas, from energy to conservation, and investigators in six continents.

Funding awarded by primary proposal subject area.

We were amazed and heartened by the high quality of the submissions to the Innovation Grants program. These proposals collectively represented significant potential impact for addressing climate change, and there were many high-quality proposals that we were unfortunately unable to fund. We’re therefore fundraising to expand the size of our funding program in the coming years, and hope that others will join us to help fill in the gaps.

For administrative reasons, the Innovation Grants were also limited to primary applicants from OECD countries (though teams could span multiple countries, including non-OECD countries). In future iterations of funding programs, we hope to drop this restriction and particularly fund applications from the Global South.

We’d like to thank Schmidt Futures and Quadrature Climate Foundation for their financial sponsorship of this program, and Future Earth for their significant administrative support.

The funded proposals are listed below (in alphabetical order):

	Accelerating Material Discovery for High-Performance Chemical Separation using AI Subhransu Maji (University of Massachusetts, Amherst); Peng Bai (University of Massachusetts, Amherst)
	Adaptive Learning Techniques for Improved Subseasonal Forecasting Steve Easterbrook (University of Toronto); Judah Cohen (AER); Lester Mackey (Microsoft Research); Soukayna Mouatadid (University of Toronto); Genevieve E Flaspohler (MIT); Paulo Orenstein (IMPA); Ernest Fraenkel (MIT)
	Detecting Flooding in Fiji’s Croplands John M Duncan (The University of Western Australia); Bryan Boruff (The University of Western Australia); Nathan Wales (The University of Western Australia); Solomoni Nagaunavou (Ministry of Agriculture, Fijian Government); Renata Varea (The University of the South Pacific); Kevin Davies (The University of Sydney); Eleanor Bruce (The University of Sydney)
	Estimate the ice volume of all glaciers in High Mountain Asia with deep learning (ICENET) Niccolo Maffezzoli (Institute of Polar Sciences - National Research Council); Eric Rignot (University of California Irvine); Carlo Barbante (Institute of Polar Science - National Research Council)
	Extracting and Discovering New Measurements from Climate Text Sources Taylor Berg-Kirkpatrick (University of California San Diego); Tom Corringham (Scripps Institution of Oceanography)
	ForestBench: Equitable Benchmarks for Monitoring, Verification and Reporting of Nature-Based Solutions with Machine Learning Dava Newman (MIT); Moises Exposito-Alonso (Carnegie Institution for Science); Lucas Czech (Carnegie Institution for Science); David Dao (ETH Zurich); Björn Lütjens (MIT); Lauren Gillespie (Stanford University); Hilary Hao (Climate Reality Project); Andrew Cottam (Restor)
	Improving Resiliency of Malian Farmers with Yield Estimation: IMPRESSYIELD Esra Erten (Istanbul Technical University); Ramazan Gokberk Cinbis (METU); Osman Baytaroglu; Traore Haoua (OKO Finance)
	Learning Power System Dynamics in the Frequency Domain Baosen Zhang (University of Washington); Weiwei Yang (Microsoft Research); Yixing Xu (Breakthrough Energy)
	Machine Learning-based Dynamic Climate Projections for Power System Planning Datasets Bri-Mathias S Hodge (University of Colorado Boulder); Aneesh Subramanian (University of California, San Diego); Claire Monteleoni (University of Colorado Boulder); Himanshu Jain (IIT Roorkee)
	Matching Structured Energy System Data for Policy Making and Advocacy using Weakly Supervised Machine Learning Xu Chu (Georgia Tech); Zane Selvans (Catalyst Cooperative)
	Mitigating Climate Change Impacts on Biodiversity via Machine Learning Powered Assessment Oisin Mac Aodha (University of Edinburgh); Scott Loarie (iNaturalist); Thomas Brooks (IUCN)
	Towards greener last-mile operations: Supporting cargo-bike logistics through optimized routing of multi-modal urban delivery fleets Maria S Astefanoaei (IT University of Copenhagen); Akash Srivastava (MIT–IBM Watson AI Lab); Kai Xu (The University of Edinburgh); Nicolas Collignon (Kale Collective); Esben Sørig (Kale Collective); Soon Myeong Yoon (Kale Collective)
	Using Machine Learning and Earth Observation Data to Identify Aquaculture Sites with High Potential for Production Intensification and Mangrove Restoration in Southeast Asia Jack Kittinger (Arizona State University); Dane Klinger (Conservation International); Pia Faustino (Thinking Machines Data Science)

^{The icons used in the table above are obtained via remixicon.com.}

The paper that launched CCAI

2022-03-08T00:00:00+00:00

In June 2019, the team that would become Climate Change AI released a 60-page paper called “Tackling Climate Change with Machine Learning.” This month, our paper was formally published in ACM Computing Surveys. In this post, we look back at the paper, the choices involved in writing it, and what we’ve learned since.

What’s in the paper

“Tackling Climate Change with Machine Learning” covers ML applications across many climate-relevant fields and sectors, using a taxonomy inspired by that of the Intergovernmental Panel on Climate Change (IPCC). Different sections address electricity systems, transportation, land use, climate science, etc. Every section is divided into areas of application (e.g., “freight routing and consolidation”), each with a brief literature review and a set of recommendations for future work. (See our interactive summary of the paper.)

An overview diagram from “Tackling Climate Change with Machine Learning” summarizing AI applications in farms and forests, one of the 13 solution domains explored by the paper.

We found several themes that cut across application areas. AI can turn raw data into useful information—for example, by using satellite imagery to automatically estimate carbon stock or risks from coastal flooding, or by parsing corporate financial reports to identify climate-related disclosures. AI can optimize complex systems—for example, by reducing the energy needed to heat and cool buildings, or by boosting the efficiency of freight transportation networks. AI can improve forecasting—for example, by predicting supply and demand to help manage electrical grids, or by forecasting crop yields. And AI can accelerate the process of scientific modeling and discovery, such as by speeding up the search for new materials like those in photovoltaic cells and batteries.

Alongside our overview of key opportunities, we discussed the overall role of machine learning in climate action. We particularly warned against techno-solutionism: machine learning, we emphasized, is not a silver bullet; it is merely one tool in the fight against climate change, and it is impactful only when used in combination with tools from other fields or sectors. Moreover, while cutting-edge machine learning is sometimes necessary, at other times simple methods suffice and should always be tried first.

Most importantly, we emphasized the need for cross-disciplinary collaboration: bringing together experts both from ML and from the relevant application domain(s), as well as the stakeholders and communities involved in or affected by a given project. Such partnerships are necessary to ensure that the right problems are being solved; that algorithms incorporate domain knowledge where possible; and that there is a feasible and responsible pathway to deployment and impact.

What’s not in the paper

Since the paper’s release, our taxonomy of key applications has mostly held up well, but in retrospect we might have tweaked some choices. For example, we neglected to discuss machine learning applications in waste management, including waste sorting and reducing emissions from landfills and wastewater. We also could have further discussed water security alongside our subsection on food security.

Additionally, many of the areas of work we discussed have grown significantly over the past few years. For example, the literature on power systems optimization has become much richer, including new lines of work using (deep) reinforcement learning. Likewise, machine learning is now being used extensively to monitor point sources of greenhouse gas emissions.

Of course, such shifts are inevitable; a paper at best represents a snapshot of the field at the time it is written. In recognition of this, Climate Change AI has started to build a community-contributed wiki, which we hope will provide a living literature review alongside other key information such as available datasets.

There are also many broader considerations that we did not delve into in the paper. For example, the bulk of our recommendations were for researchers and practitioners rather than policymakers. Since the release of “Tackling Climate Change with ML,” Climate Change AI has released a separate report—co-authored with the Centre for AI & Climate for the Global Partnership on AI (GPAI)—which offers recommendations for governments to facilitate work at the intersection of ML and climate change. The report includes strategies for capacity-building, opportunities for catalytic funding, and key areas for international cooperation.

We also did not discuss in detail the many ways in which machine learning is making climate change worse—for example, the widespread applications of AI to accelerate fossil fuel extraction, the emissions associated with hardware and computation, and the broader effects of AI-based technologies on societal consumption patterns. “AI for Good” should mean more than just adding beneficial applications on top of business as usual. Several of us have since co-authored another paper that dives more deeply into this topic, calling for climate-related impact assessment across machine learning as a whole.

How the paper came together

The paper was motivated by the recognition that there were significant opportunities at the intersection of ML and climate change, but that relatively few of them had gained widespread attention. On the machine learning side, many researchers and practitioners were anxious about climate change, but didn’t know how their skills could be useful, or what working on climate change even meant. Meanwhile, in climate change-relevant fields like energy, climate science, and disaster response, experts often lacked the experience in machine learning to pinpoint how ML tools could (or should) be used.

Our goal in writing the paper was to mitigate these problems by:

identifying areas of climate action where machine learning could be highly impactful;
highlighting existing communities and work at the intersection of machine learning and climate change; and
providing a call to action and a roadmap for diverse researchers and practitioners to work in this space.

The paper had to be assembled in about six months, as we were aiming to launch it alongside the climate change workshop we were organizing at the 2019 International Conference on Machine Learning (ICML). This timeframe was tight, but the work was parallelized across different sections, each spearheaded by a section lead with relevant expertise. The timeline was roughly:

Months 1-2: Literature review and consultation with a broad set of experts
Month 3: Identifying key recommendations
Month 4: First drafts of sections
Months 5-6: Integration and editing

As the sections took shape, it was important that the paper held together as a coherent whole. To that end, we set common standards for organization, level of detail, and writing style; we held weekly meetings to exchange feedback; and we ran the final paper through several rounds of revision, both internally and with external stakeholders. To help the reader navigate, we added summary tables, one linking application areas with specific subfields of machine learning and the other linking application areas with cross-cutting themes on the role of ML.

We also added tags that allowed the reader to jump to especially high-leverage applications. However, we decided against ranking areas of application, feeling that any attempt to quantify their relative impacts would require too much speculation. In addition, we feared that a top-ranked solution could come across like a silver bullet, when in fact climate action requires working on many important things simultaneously. We also created tags for applications with particularly long time horizons for impact (such as afforestation), or where the impacts were uncertain or potentially detrimental (such as autonomous vehicles).

Looking ahead

In the wake of “Tackling Climate Change with Machine Learning,” we have been glad to see climate change increasingly cited as a flagship application of “AI for Good.” We are happy that our paper has been used as a resource in many different ways—by governments developing grants programs and digitalization strategies, by companies identifying opportunities, and by individuals picking jobs or areas of study.

But cultivating impactful work requires much more than just a guide to key applications and a call to action. That’s why we at Climate Change AI have been working on activities aimed at alleviating bottlenecks across the space. Many of our efforts target bottlenecks identified in the paper, such as the difficulty in establishing cross-disciplinary teams and existing gaps in resources and data. There is much to be done, and we hope that researchers and practitioners from machine learning and other disciplines will join us in our continued work to help tackle climate change.

CCAI Core Team Profile: Evan Sherwin

2022-02-07T00:00:00+00:00

The CCAI blog occasionally features profiles of core team members. This week, we check in with methane researcher and CCAI core team member Evan Sherwin.

Tell us a little about yourself.

I study the future of hydrocarbon fuels in a net-zero-emission energy system, focusing on methane emissions from oil and gas and synthetic hydrocarbons in my postdoctoral work at Stanford University in California. With a PhD in Engineering and Public Policy and an MS in Machine Learning from Carnegie Mellon University in Pennsylvania, I use engineering insight to produce data-driven analyses to inform government and industry emissions reduction efforts. I founded and chair the Methane Emissions Technology Alliance international webinar series and was a founding member of Climate Change AI.

What role(s) do you play within CCAI?

As the Chair of the Programs Committee and a member of the Board of Directors of Climate Change AI, I support teams producing our various programs, including conference workshops, webinars, happy hours, community events, grants, and our summer school. My primary responsibility is mentorship, community-building, and guiding the future of the organization. I also provide hands-on support as needed. I co-led the 2020 Tackling Climate Change with AI NeurIPS workshop.

How did you get involved with CCAI?

I was an author of the Tackling Climate Change with Machine Learning report, and helped found Climate Change AI to help achieve the vision we outlined.

What are some common paths people take to doing the kind of work you do?

The road to a position like mine generally includes its twists and turns. A PhD is generally necessary, but it often pays to work in industry or consulting for a few years beforehand. I spent three years in quantitative energy policy consulting after completing my undergraduate degrees in physics and applied mathematics at University of California, Berkeley. This helped me find what I felt was an impactful path in graduate school and beyond.

What motivates you to work at the intersection of AI and climate?

Remote sensing is a powerful tool to quickly find and flag large methane emissions. Much of my work focuses on characterizing the policy and industrial implications of the rapid development of aerial and satellite-based methane remote sensing systems. Also, a lot of machine learning is just statistics, and sometimes plain old statistics is the key to answering really important questions.

Who or what inspires you?

The powerful response that CCAI continues to receive across disciplines, sectors, and continents has been overwhelming. The knowledge that I can help build a community of brilliant problem-solvers to tackle climate change is an inspiration.

Don’t be afraid to pick up new disciplines as you go along. Also, however much statistics you’ve learned already, learn more.

The EPA’s new methane reduction proposal hints at help from AI

2022-01-19T00:00:00+00:00

On November 2nd, the Environmental Protection Agency (EPA) put forward new regulations to help limit methane pollution. As I was browsing the headlines about this announcement, one in particular caught my attention:

Slate reporting on the EPA’s new climate rule.

What mistake is the current administration trying to remedy?

The answer turns out to make a big difference for climate change—and also represents an excellent opportunity for AI to help.

Methane on the rise

Let’s start with why methane is such a concern in the first place. Most people concerned about climate change focus on carbon dioxide from burning fossil fuels, but methane from multiple sources, including leaks, poses just as serious a risk. Methane can escape into the atmosphere during oil processing, storage and distribution—so much so that a recent study estimates that in some regions, approximately 3.7% of all natural gas extracted is lost. Additionally, although methane does not last as long in the atmosphere as CO₂, it traps more energy, so a small reduction in methane provides similar benefits to reducing larger amounts of CO₂. The EPA estimates that its new rules will eliminate 41 million tons of methane emissions from 2023 to 2035, which is comparable to 920 million metric tons of CO₂.

That reduction is especially important in light of the very concerning upward trend in atmospheric methane concentrations since 1987. I graphed out measurements from two of NOAA’s global monitoring laboratories, one in Mauna Loa, Hawaii (MLO) and one in Barrow, Alaska (BRW), as well as the average of the two locations’s measurements (AVG). As you can see in the graph below, methane emissions have increased over 10% in that time period—a dangerous trend that these new regulations aim to reverse.

Increasing methane emissions over time.

No more grandfather clauses

The biggest issue that the EPA’s new rules address is that the methane emissions of over 300,000 facilities are not currently monitored. Under previous administrations, regulations handed down from the EPA were not applied to then-existing well sites. With the grandfather clauses removed in the latest regulation proposal, companies will now be required to monitor those facilities once per quarter on the ground, while remote screening can take place every other month.

Many smaller well sites remain uncovered by this proposal. Still, the EPA estimates that by requiring all wells that emit at least 3 tons of methane per year to be monitored quarterly, the new rules will target 86% of all methane leaks from well sites.

Monitoring more is not the only change in the new proposal. The EPA is also pursuing a number of mandatory upgrades to many oil and gas sites and systems. These include:

requiring zero emissions from pneumatic controllers, which account for almost 30% of all methane emissions from oil and gas system, and from pneumatic pumps;
eliminating venting and flaring of gas from wells with no sales line;
expanding the types of facilities covered by their rules; and
minimizing methane emissions from unloading liquids.

All of these proposals aim to reduce methane leaks and reduce the amount of methane in the atmosphere.

AI will both affect and be affected by the changes

These new rules will interact with AI in two key ways. First, they will directly affect several current AI projects. A number of machine learning models aim to predict methane emissions, such as work from Arvind Ravikumar’s group that uses probabilistic classification to identify high-emitting methane sites. Assuming the EPA’s new rules are effectively enforced, this work and other projects like it will likely need to be updated to account for shifts in emissions patterns.

The effects can also run in the other direction, though, with AI applications helping to enforce the new rules. One group at the University of Chicago is developing tools to quickly analyze satellite images to identify “hotspots” of greenhouse gas emissions, including methane. Another example is hyperspectral imaging (imaging sites at many different wavelengths), one of the ways methane emissions can be monitored. Inconveniently, methane’s spectral signature is similar to many other hydrocarbons’, so BS Manjunath’s group at UC Santa Barbara is using deep learning to help identify methane leaks in hyperspectral imaging based on the speed and spread of previous leaks. Machine learning algorithms are also used in commercial products that help monitor flare performance, catch equipment failure before it happens, and predict buildings’ energy needs. While some of these technologies may lie outside of the purview of the EPA’s new methane rulings, all of them will help in the mission to decrease waste and emissions.

The EPA recognizes the advances in methane detection but gestures only vaguely at their applications

The ways AI can help with methane detection and leak reduction have not gone unnoticed by the EPA. In their summary of the new proposal, they dedicate a call-out panel to Advanced Methane Detection Technology, and propose giving well owners flexibility in which detection approach to choose.

However, the EPA is being a little stricter with the standards for AI-driven approaches. Owners who choose these new technologies will be required to self-survey every two months (rather than every three), and also to supplement their advanced surveys annually with more standard optical gas imaging. Given the nascent state of advanced technologies, it doesn’t seem surprising that the EPA would be hesitant to commit to trusting them right away, but this open-mindedness suggests a promising future for AI-driven techniques in methane detection and reduction.

The EPA is currently seeking information and input from experts on this proposal, which the agency will incorporate into a supplement in 2022. If you would like to provide comments, you can do so at the EPA’s website (until January 31), and if you would like to read the proposal in full you can do so here.

Climate Change AI @ COP26

2022-01-12T00:00:00+00:00

“The future is watching”: A dispatch from COP26

“No more blah blah blah!” “COP26 act now!”

The closer you get to the Scottish Event Campus, the more clearly you can see the protest signs and the louder you can hear people chanting. Hundreds of protestors, young and old, gather daily in front of the security gates of COP26 to remind a long queue of delegates what’s at stake here in Glasgow.

In the morning, the queue to enter COP sometimes gets so long that delegates have to wait for hours. It’s enough time to read all the protest signs and reflect on them: “The future is watching.” “If you fail, we will never forgive you.”

The world has been waiting for this Climate Change conference for more than two years. 2020 has been the hottest on record. Global warming has already reached 1.2°C above pre-industrial levels, and sea levels are rising. For many delegates in line, it is clear that this is our last decade to act.

Weirdly, the moment you pass through security and enter COP26, the sounds of protesters vanish. Instead, you find yourself in a conference that feels almost too normal. Delegates from all around the world run from one meeting to another. Countries have pavilions where they host events, provide snacks, and hand out souvenirs. There are restaurants, food courts, art exhibitions, and a large plenary hall where arguably the world’s most important negotiation will take place over two weeks.

Fortunately, you never get too used to this feeling of normality: waiting in line the next morning reminds you of the weight on your shoulders when you enter. You feel the world’s gaze.

– David

CCAI and COP

Since 2017, I’ve had the privilege of attending the COP repeatedly, first as an NGO observer and later as a United Nations observer. I’ve seen delegates oversell artificial intelligence and disruptive technologies to mislead politicians into imagining easy solutions to the climate catastrophe. Corporations also spotted their chance to greenwash: they started advertising disruptive technologies, at COP and beyond, as an excuse to not phase out their emissions.

Climate Change AI first held an event at COP25 in Madrid, shortly after the organization was founded in 2019. The discussions of AI at CCAI’s event brought a then-rare emphasis on science, realistic pathways to impact, and ethical implications of technology. What I saw convinced me to join the organization shortly afterwards.

At COP26 this past November, CCAI continued serving the same role: with its partners, it organized several events that brought scientific input to policy discussions and distinguished hype from meaningful action. COP26 also provided the stage for launching a report that CCAI recently co-authored for the Global Partnership on AI (GPAI), which details policy recommendations to foster climate-relevant AI applications.

– David

COP26 Event 1: “AI for Climate Action”

CCAI’s first event, “AI for Climate Action,” took place at the German pavilion. The event was jointly organized with the Centre for AI and Climate (CAIC) and the German Federal Ministry for the Environment, Nature Conservation, and Nuclear Safety.

While many pavilions held in-person events just like before the pandemic, the German pavilion was designed around online streaming: it was held in what looked like a fish tank that could be watched both from the hallway outside and online. The pavilion’s tech crew and a producer managed to make our mix of in-person and virtual speakers, moderators, and audience—which could have failed miserably—an incredibly smooth event.

The German pavilion’s “fish bowl” setup

The event kicked off with three presentations of use cases where AI is being applied to address climate change. Dr. Arshad Mansoor, the CEO of the Electric Power Research Institute (EPRI), spoke about how EPRI leverages ML in power systems research and practice. We then heard from Edoardo Nemni of the United Nations Satellite Centre (UNOSAT) about AI’s role for humanitarian relief and emergency response to floods, and from Dr. Irene Sturm (Digital Rail for Germany, Deutsche Bahn) how reinforcement learning can be used for planning and traffic management in future rail systems.

In the second part of the event, six panelists discussed governance approaches to AI in the context of climate action and the COP. Pete Clutton-Brock (CAIC) introduced the opportunities and policy approaches detailed in the recent report by CCAI, CAIC, and the Global Partnership on AI (GPAI). Dr. Marta Kwiatkowska (University of Oxford), who also contributed to the report, added the caveat that it is also important to understand and assess the negative impacts of the technologies on the climate. She especially emphasized the important role of international governance initiatives.

One avenue for governments to foster AI for climate applications is by providing open data. Dr. Cristóbal de la Maza (Chilean Superintendency of Environment) discussed Chile’s recently-launched Environmental Intelligence Strategy (Inteligencia Ambiental), which includes using AI to process large amounts of openly available environmental data collected in the country. Dr. Catherine Nakalembe (University of Maryland, NASA Harvest) then shared mirror-image examples from AI applied to food security, in which data availability acted as a bottleneck in her work to forecast crop health and yields from satellite images.

Another way governments can help is by providing data annotations, benchmarks, and simulators. Erika Gupta (US Department of Energy [DOE]) described the DOE’s efforts to spur energy efficiency innovation in the building sector by creating benchmark datasets hand-in-hand with standardized building models.

When incentivizing new AI projects, governments also have to avoid promoting applications that counteract climate goals. Daniel Schmitt (German Federal Ministry for the Environment, Nature Conservation and Nuclear Safety) noted that this is a significant challenge: the technology is moving so fast that by the time a government develops a framework for what funding targets support, it might be already outdated. “It is really tough to address these issues from a governmental funding perspective,” he said. Strong alliances between national governments, as well as between government and different stakeholder groups within society, can help to address this challenge.

– Lynn

COP26 Event 2: “A Clearer Picture: Towards Radical Transparency in Measurement, Reporting and Verification of Climate Action with AI”

Our second event discussed the role of AI in measurement, reporting and verification of climate goals (commonly abbreviated as MRV). The discussion was the closing event hosted by the UN Climate Change Pavilion.

CCAI’s closing event on MRV at the UN Climate Change Pavilion

MRV is a key building block for successfully implementing and tracking “nationally determined contributions”—the climate-related targets that each signatory nation to the Paris Agreement commits to. Together with ClimateTRACE and CAIC, CCAI organized a panel discussion to inform decision-makers about how AI can increase the transparency and accountability provided by MRV efforts.

Satellites capture 80TB of data every day. Global alliances such as ClimateTRACE can now train machine learning models on this data to generate estimates of emissions from all sectors, offering policy-makers an independent alternative to national self-reports. “We are no climate cops, but you can think of us rather as your friendly climate neighborhood watch,” said Matt Gray, co-CEO of TransitionZero, a founding ClimateTRACE member (and “best pathway to impact” winner at CCAI’s NeurIPS 2020 workshop).

Dr. Marco Schletz from the University of North Carolina emphasized how AI can work with other technologies such as blockchain to help achieve the goals of the Paris Agreement. Dr. Dava Newman, director of the MIT Media Lab, went one step further: she pointed out the importance not just of leveraging satellite-based data and technology, but also of working across all disciplines and fields—including social sciences and art—to ensure that the methodologies and predictions that are developed are fair, inclusive, and interactive.

Dr. Alejandro Paredes Trapero from FSC Indigenous Foundation further explained the important role that local communities and Indigenous People can play in co-designing AI-based monitoring solutions.

The XPrize foundation, represented by Luiza Barguil, couldn’t agree more with Alejandro. Luiza showcased how teams competing for the XPrize for biodiversity monitoring are already benefiting from the guidance of their indigenous jury members.

CCAI believes that artificial intelligence is a powerful tool that can help tackle climate change—but it is by no means a silver bullet. That message clearly resonated with the other COP26 delegates. (“We couldn’t have imagined a better closing event”, one member of the UNFCCC secretariat told us afterwards.)

We are thankful for all the support we received to make these events happen, and especially grateful for the opportunity to inform decision-makers at the highest levels.

– David

Learning to Control Buildings Like an Engineer

2021-12-21T00:00:00+00:00

Buildings are a big deal for climate change. They contribute to 40% of global energy use, with a large portion of that energy going to heating, cooling, ventilation, and air-conditioning (HVAC).

AI can play a major role in reducing that energy use. Switching to AI-based control can reduce energy use and greenhouse gas emissions by 13-28%. In fact, if all buildings were operated by advanced AI systems, we could reduce the global energy/emission bill by at least 4%!

Nonetheless, in most 21st-century buildings, HVAC systems are still controlled by “dumb” rule-based controllers (RBC) that are designed and tuned by hand. Why are we not already using advanced AI to improve the energy efficiency of buildings at scale?

The problem is that switching to AI control isn’t as easy as replacing your smartphone or buying a new car. Buildings are made to last, and even worse, each building has a unique layout, infrastructure, location, and set of user profiles. All of this makes modernization very costly.

In this post, I focus on the problem of designing advanced control methods for HVAC systems in a cost-efficient way. In particular, I discuss how to augment data-driven deep learning with domain-aware physics-based control using a hybrid method called differentiable predictive control (DPC). DPC essentially allows a data-driven system to incorporate an approximate physics simulator into its training. This enables it to learn a reliable, interpretable model with far less training data than previous data-driven methods.

Which way to the promised land?

Currently, engineers design smart HVAC control systems using two competing approaches, model-based and data-driven methods.

In model-based approaches, the advanced controller is designed by engineers from known physical laws. If done diligently, model-based methods, the most popular of which is called model predictive control (MPC), can provide outstanding performance with safety guarantees.

In an ideal world, then, engineers would simply deploy an MPC system for each new building. Unfortunately, each new building’s model must be designed by a human expert, making this method tedious and costly in practice. These controllers are also expensive to maintain due to complex optimization routines that need to run in real-time.

The main alternative is data-driven approaches such as deep reinforcement learning (RL), which promise automated controller design via the powers of machine learning. Unfortunately, purely data-driven approaches do not provide safety and performance guarantees. They also must learn the best control policies without any knowledge of how buildings and temperatures work, which often requires prohibitively large datasets.

We can deal with this conundrum by fusion. Not nuclear fusion (though it would be cool to have a portable tokamak in the garage!), but rather by fusing model-based and data-driven AI methods for optimal control design.

In a recent series of papers, my PNNL colleagues and I described a new method called differentiable predictive control (DPC) for learning model-based HVAC control policies. The core idea is to create a neural network that approximates a building’s thermal dynamics. This model can then be incorporated directly into a deep reinforcement learning controller to help train the overall system.

Our method has two steps, illustrated in the figure below. We start from a real-world dataset of climate conditions, HVAC settings, measured building temperatures, and so forth for a given building. With that data, we first train a data-driven predictive model of the building dynamics, designed to prefer models that exhibit known physics-based behaviors. Second, we use the learned model, known performance metrics, and physics-derived constraints and optimization penalties to optimize the control policy in a domain-aware way.

As a result, we obtain a highly effective method that combines the benefits of both worlds. Thanks to machine learning, it can quickly adapt to new scenarios. And thanks to inductive biases drawn from physics, it can do so using small datasets while satisfying operational constraints.

Differentiable Predictive Control (DPC) methodology. Step 1: Learn a physics-guided neural model of buildings’ thermal behavior from the gold-standard time series dataset produced by a building physics emulator. Step 2: Learn neural control policies by backpropagation of the MPC gradients through the neural system model.

Step 1: Physics-Constrained Neural Surrogate Model of the Building Dynamics

In the first step, we automatically learn the thermal behavior of a given building via a physics-constrained neural state-space model (SSM). The model has a set of state variables, representing the temperatures of the building walls and indoor environment, and a set of control inputs, which are manipulable variables of the HVAC system such as mass flow rates and supply temperatures. We also consider a set of measurable disturbances representing the environmental factors such as ambient temperature and solar irradiation. Then, given a time-series dataset, we train this SSM model to predict the future evolution of the indoor temperatures that satisfy a set of high-level constraints inspired by laws of thermodynamics.

We structure the computational graphs of these models to mimic the dynamics of physical buildings. We decompose the whole system into subsystems, or blocks, where each of the blocks represents a different part of the physical system—e.g., the building envelope, HVAC system, or influence of weather conditions. To make the model easily transferable to different types of buildings and operational conditions, we represent each individual block by a deep neural networks.

To improve accuracy and generalization, we guide the neural networks by penalizing violations of certain thermodynamic laws. One key physics insight is that a building represents a dissipative system evolving to achieve thermal equilibrium with the environment. This inspired us to enforce constraints on the model that can be interpreted as promoting the dissipativeness of an abstract dynamical system. Furthermore, using an optimization strategy known as the penalty method, we can further nudge our model to keep its states within physically realistic bounds.

In a case study, we demonstrate the learning of a thermal dynamics model for a real-world building, given a limited amount of measurement data. For our experiments, we did not have access to the full real-world dataset, so we instead used a simulated dataset generated by a pre-existing physics emulator, which models a building’s thermal dynamics reasonably accurately.) We considered a commercial office building with a typical setup, including air handling units, 20 thermal zones, a natural gas boiler, hot water reheat coils, and a central chiller plant. When we trained and validated our system on just 20 days of data (with less than 100 data points per day), it predicted the next 10 days’ thermal behavior with less than 50% of the error rate of previous data-driven methods.

Trajectories of the learned (blue) and ground truth (red) multi-zone building thermal dynamics. Grayed areas denote train, development, and test sets, respectively.

Step 2: Differentiable Predictive Control

Once we have an accurate model of the building’s thermal dynamics, we can proceed to the second step, model-based policy optimization.

In conventional RL approaches, the system must learn everything from scratch. It has no model of the building and or of thermodynamics; all it has is measurement data and some “reward signal” indicating how well a given set of control decisions performed. Even if the reward signal comes from a physics model, rather than the actual building, learning a control policy in this way is still extremely inefficient.

We could mitigate this problem if we could prime the RL system with a model of the building’s thermal dynamics. That way, the only thing to learn by reinforcement would be the control policy; the building model would already know how the environment behaves. But to leverage an embedded building simulation, each round of training would have to try a control policy, run the building simulation, calculate the reward/error for the results, and adjust the neural network parameters for the control policy accordingly. With a conventional physics-based simulator, there is no way to propagate the reward/error signal back to determine how to adjust the control parameters.

This is where our neural surrogate model comes in. Because our approximate model of the building’s physics is itself a neural network, it is fully differentiable, meaning we can leverage the powerful automatic differentiation tools underlying all practical successes of deep learning. This design trick allows us to backpropagate the control performance metrics (rewards and constraint violation penalties) through the system dynamics model to optimize the policy parameters directly. Thus, the reinforcement learning only has to tune the control parameters; it does not need to learn the environment dynamics. As a result, the DPC method is much more sample-efficient (and interpretable!).

Now we are well equipped to deal with the challenging problem of controlling the heating system of a multi-zone building. We train the DPC policies by sampling expected operational conditions from the learned surrogate model and letting training work out optimal policies for those scenarios. The control performance of the trained DPC policy is shown below.

Closed-loop control performance over time of a DPC control policy evaluated on the learned building model. On the left-hand side, we plot in color the temperature in individual zones that ought to satisfy the given comfort constraints shown in dashed black. On the right-hand side, we show the corresponding mass flow rates of the heating system delivered to each of the zones.

Finally, we validate the DPC method’s performance in a case study using a physics-based simulator of the thermal dynamics in a multi-zone building. Our results indicate that DPC can handle economic objectives subject to dynamic constraints imposed on nonlinear system dynamics with multiple inputs and outputs.

Conclusion

Energy-efficient building controls could be revolutionized by reliable, data-driven methods that are cost-effective in terms of computational demands, data collection, and domain expertise. Our two-step DPC learning method is a large step toward this goal.

The DPC approach does not require the large time investments by domain experts and extensive computational resources demanded by physics-based emulator models. Based on only ten days’ measurements, we significantly improved on prior state-of-the-art results for a modeling task using a large-scale office building dataset.

Additionally, in a simulation case study using a multi-zone building thermal dynamics model, we showed that it is possible to learn control policies for constrained optimal control problems with a large number of states over long prediction horizons.

Based on the presented features, we believe that the proposed DPC method has long-term potential for research and practical applications. DPC is particularly appealing for scenarios that require fast development and low-cost deployment and maintenance on hardware with limited computational resources. More fundamental and practical work needs to be done to further verify the real-world performance, but we are optimistic about the promise of these fusion-based methods.

A datathon to empower women in climate change and data science

2021-12-13T00:00:00+00:00

The 5th Annual Women in Data Science (WiDS) Datathon will launch in January in partnership with Climate Change AI. WiDS is an initiative started by Stanford University that aims to inspire and educate data scientists worldwide, regardless of gender, and to support women in the field.

As in data science, women are underrepresented in the scientific fields studying climate change. For example, a 2018 survey of 100 female authors of the Intergovernmental Panel on Climate Change (IPCC) reported that many of these women, despite their exceptional expertise, felt poorly represented and heard. In 2020, the fraction of IPCC authors who were women still stood at less than a third, up from a mere 10% in 1990. The slow progress and continuing low numbers speak to a joint need to increase gender representation within both data science and climate change, including at the intersection of these fields.

The 2022 WiDS Datathon will work toward this goal by offering an opportunity for anyone—including and especially women— to hone their data science skills on a climate challenge. Specifically, the Datathon will focus on the role of buildings in reducing global greenhouse gas emissions. Decarbonizing building construction and operations is a necessary step towards achieving climate neutrality: as of 2020, the buildings sector accounts for 37% of global energy- and process-related CO2 emissions, and that share is only expected to grow. AI can provide useful tools for achieving low-carbon buildings.

Climate action in the building sector also has multiple relevant linkages with women’s well-being and empowerment. For example, the burden of heat waves is disproportionately carried by women, and women suffer most from the health effects of indoor pollution due to traditional cooking fuels. Hence, strategies that improve thermal comfort or access to clean, low-carbon energy can indirectly support gender equity.

The Datathon task will consist of predicting the energy use of buildings based on various predictors including climate and building characteristics. Given a dataset and resources such as tutorials, participants will be challenged to use their creativity and data science skills to build, test, and explore solutions. Real-world use cases of predicting building energy use include inferring missing data across the building stock—an essential task, given the scarcity of granular energy data—and forecasting future consumption in different scenarios.

The WiDS Datathon will run from early January to late February 2022 on Kaggle, an online data science community and competition platform. The dataset and challenge will be accessible to both beginners and experienced participants of all genders. You can learn more about the context of the datathon and how to get started on the WiDS announcement post.

Last year, the WiDS Datathon saw record-breaking participation, with nearly 4,000 registrants and more than 24,000 entries. We encourage you to consider participating in this year’s edition. You can sign up here!

GANs for Dynamic Spatio-Temporal Patterns

2021-12-02T00:00:00+00:00

As our climate changes, tools for understanding and modeling our planet are becoming more important than ever. Simulating Earth systems and geophysical processes is crucial for a myriad of tasks, from modeling coastal areas for flood prevention to understanding climate dynamics through turbulent flow modeling. But granular models of our planet are difficult to tune and often rely on computationally expensive numerical simulations.

One promising approach is to apply machine learning, which can model complex, non-linear data at scale. ML is especially appealing given the increasing availability of near real-time Earth observation data, collected by fleets of satellites and pervasive mobile sensing infrastructure on the ground. These rich data sources, together with powerful neural network models, could help us simulate ocean dynamics and forecast extreme weather events.

However, recent research shows that traditional deep learning approaches might not be equipped to handle the intricate spatial, temporal, and spatio-temporal complexities inherent to these data. Earth systems are governed by physical laws manifesting over space and time, and neural networks have no built-in intuition for those. To overcome these issues, as the authors note, it is crucial to augment deep learning approaches with domain expertise from the geospatial sciences.

This call-to-action catalyzed a growing interest in physics-informed deep learning, in which (geo)-physical constraints are integrated directly into models. My colleagues and I recently proposed such a method for modeling spatial dependencies: we give the neural networks additional training on auxiliary tasks that involve predicting measures of spatial autocorrelation. The intuition is that by providing models with explicit information on how spatial dependencies manifest in the data, we can nudge them to learn these processes during training.

In our latest paper, we expand this idea to the spatio-temporal domain. But how can we measure spatio-temporal autocorrelation? A look into geography can help! This field has a long tradition of dealing with the complexities of geospatial data.

Geographers have developed metrics like the Moran’s I statistic to empirically assess spatial autocorrelation, which describes an observation’s dependency on its local spatial neighborhood. We can turn Moran’s I into a new measure for spatio-temporal autocorrelation by making one small tweak: Moran’s I averages differences between grid cells over space. We replace those computations with averages over space and time. We refer to this new measure of spatio-temporal association as SPATE.

Specifically, we devise three variants of SPATE that compute these space-time expectations in different ways: (1) without temporal weighting and with access to future time-steps; (2) with temporal weighting and with access to future time-steps; and (3) with temporal weighting and without access to future time-steps.

We then deploy SPATE in a generative adversarial network (GAN) application. GANs learn the data-generating process of the training data and can generate high-quality synthetic data samples. My team and I propose SPATE-GAN, a GAN augmented with our new SPATE metric. We accomplish this by setting up the loss function during training to reward the model for generating synthetic data that reproduces observed spatio-temporal dynamics.

We test SPATE-GAN on three relevant real-world problems: spatio-temporal point processes (used for, e.g., modeling disease spread), simulating surface temperature, and simulating turbulent flows (e.g., atmospheric and ocean currents). All these tasks exhibit intricate spatio-temporal patterns governed by complex physical laws. And compared to previous methods, SPATE-GAN can indeed improve performance, generating more realistic synthetic data.

Global surface temperature patterns over 10 time steps. Displayed are real observations and generated, synthetic samples from SPATE-GAN and other benchmark models.

There is still a long way to go until, for example, deep learning methods outperform large numerical climate simulators. Still, studies like ours and many more from the physics-informed deep learning community highlight the potential of deep learning for modeling our planet at scale.

These studies are also humbling reminders, however, that we cannot just throw data into multi-billion parameter neural networks and hope for a good outcome. We must take a more balanced approach, combining expertise from the machine learning community with insights from relevant application domains.

Startups in the Climate Space: An Interview with Dr. Lauren Kuntz of Gaiascope

2021-11-12T00:00:00+00:00

This is a conversation between two startup founders about hurdles we’ve faced. We share our advice for startups trying to save the climate.

My interviewee, Lauren Kuntz, is CEO and co-founder of Gaiascope, a Y-Combinator funded startup building software that forecasts price dynamics in the US electric power grid.

My own startup experience comes from serving as CTO and co-founder of the TRASH app for video editing, acquired in 2020 by Visual Supply Company (VSCO).

Lauren and I are also both CCAI core team members and co-creators of CCAI’s Machine Learning + Climate Intro Course.

Patterson: Lauren, you got a PhD in Climate Science before starting a company to forecast renewable energy supply, weather, and grid demand. I got a PhD in machine learning before I started an AI video editing app. We were both arguably well prepared to be founders. But I felt that I had to take giant risks and leaps of faith to succeed as a founder! How much did your education prepare you to run your climate-impacting startup?

Kuntz: I firmly believe that you never are fully prepared to do something; you become prepared by doing it. That said, a ton of skills transferred from my education into start-up life. In particular, my PhD research helped me become comfortable not knowing an answer out of the gate and trusting I’d figure it out as I explored further. Ironically, my education helped me less in the sense of “Now I know what I’m doing”—I never fully feel like I know what I’m doing!—and more in the sense that it taught me how much I don’t know and to be OK with that. You’re going to be uncomfortable and have to take risks and leaps of faith as a founder, and that’s OK! What I’ve found helps the most is building up your problem-solving chops and learning to live with the discomfort of uncertainty.

Patterson: When I started TRASH, I thought, “Yes! Finally I get to be a genius AI inventor.” But then the ML turned out to be only about 10% of my product/company. I had to become a mobile app developer, a backend engineer, a PM, a growth marketer, a fundraiser, and the best uses of my product were still a surprise to me. In a climate startup, it seems even less likely that the fun experimentation part will lead to success. Lauren, how have you discovered what you didn’t know you didn’t know?

Kuntz: I made so many classic mistakes about assuming I understood the problem from the get-go, when really it was so much more complicated. It can be easy to simplify a problem until I can write a nice equation or algorithm to cover it, but the real world is incredibly messy! There’s generally a reason the problem you’re going after still exists. For us—trying to understand wholesale electricity markets—the reason they’re so opaque is they’re so complex. We couldn’t ignore that complexity or we wouldn’t be creating anything meaningful. It is so hard to accept all the complexity of reality, but what really helped us was being on the ground and using our own product. We use our own forecasts and algorithms, so if there’s a simplification or anything that falls short, we feel its impact on our performance. That is a massive push to do better and learn quickly.

Patterson: My startup was in the social media and entertainment space, so compared to the climate space, my success or failure seemed lower-stress. Still, I put a lot of thought into designing an app that would benefit users’ mental health and promote civil society. How can readers entering the climate space convince themselves that their startups are solving a problem that makes the climate crisis better?

Kuntz: Ideally, if you’re working on a climate problem it should be a very clear argument for why it helps. If you find yourself doing mental gymnastics to articulate the impact, you might want to step back and ask why you’re working on it. That said, the impact doesn’t have to be immediate, nor does it have to be trivial to translate into a carbon savings/reduction. I like to think 20-30 years into the future, where the world will need to be deeply decarbonized. Is your innovation one that may not work, but if it does, it will play an integral role in that future decarbonized world? For instance, a breakthrough in cheap batteries or nuclear fusion would make decarbonization massively easier. Alternatively, is this something where a decarbonized world depends on having figured it out? That’s where my work falls: to have a decarbonized grid operate efficiently, you need to have a great forward-looking model of expected prices so assets and loads can be distributed effectively. The best way I know to evaluate impact is to think about these forest-level questions: what does the future world need to look like to solve the climate crisis, and what piece of the puzzle am I working on?

Patterson: On a practical note, I found hiring and developing a strong team to be surprisingly difficult. “Save the World” is perhaps the best mission statement a team could hope for. What other strategies have you used to light the fire in your team?

Kuntz: More than anything else, I really value having a strong team. I’m incredibly grateful that at Gaiascope, I get to work with absolute rock stars. They’re not just amazing at what they do, but wonderful people to work with, learn from, and be inspired by. I’ve found that I actually don’t have to do a ton of fire-lighting, as the team is intrinsically motivated and it’s a positive feedback loop. More of my job is making sure they feel empowered to own their role, equipped to take it on, and secure in an understanding that if anything goes awry I’ll be there to help. That relies partly on making sure they feel like a team member and appreciated as a complete person with a life and values beyond work. If you’ve hired the right people, it’s not a game of motivation so much as making sure everyone is supported to go off and do amazing things.

Patterson: Fundraising gets a lot of attention in startup how-tos. For a climate startup, founders need to promise economic triumph and a carbon coup. How have you inspired investors with your net-zero future vision? And how do you turn investors into evangelists for your business and climate visions both?

Kuntz: My cofounder is an absolute queen for taking on the bulk of the fundraising, so she gets the credit for inspiring investors. As a company, our two top values are planetary balance and profitability. We believe that if you can make carbon reduction the most profitable option and show people how to make money on it, you’ll accelerate a net-zero future. For that reason, our business and climate visions are inextricably linked—we’re not a business with a climate benefit tacked on, and we’re not a climate impact group with a business tacked on. We are very much both, and I think that helps tell the story to investors. You have to believe in your vision to convince others of it—and if you don’t have faith in both the business and climate components, you probably should spend time convincing yourself.

Patterson: Starting a company is intimidating. Even more so when your goal is to get a $1B valuation or change the behaviors of hundreds of millions of people. Many people (you and I included) turn to incubators to help us get started. Do you think incubators like Y-Combinator are set up to help newcomers make a climate impact?

Kuntz: There are many incubators for many different focuses. Some definitely focus on business and impact, while others focus more narrowly on business. For us, Y-Combinator was great at helping us think through building a great business, and we already had convinced ourselves of the impact of the technology, so it was a great fit for what we needed. For anyone considering an incubator, I’d suggest they articulate what they are hoping to get out of the program—is it business development support, climate impact support, fundraising support? Once you know what support you’re looking for, it’s easier to determine which incubator can best provide that.

Patterson: Someday soon, our readers will find themselves at the helm of a company that is making money and changing the human relationship with carbon emissions. How do they evaluate their conflicts of interest—profits vs. climate mission? How are you doing that at Gaiascope?

Kuntz: Honestly, I think it is not helpful to portray these two as being in conflict. We work at the intersection of both by trying to figure out what technologies can help make green energy more profitable. When we do run into questions, we rely on those core values of planetary balance and profitability to guide us. We check in regularly to make sure we’re living up to them, and we also have worked to instill those values in our team and empowered our team to speak up when they feel we are falling short. Accountability is key, and building in that internal accountability from day 1 is a must.

For more insights on how to build your decarbonizing business, please reach out to us on our new CCAI Community Platform and/or subscribe to the CCAI Newsletter!

CCAI’s comments on the EU’s proposed Harmonized Rules on AI

2021-11-02T00:00:00+00:00

In April this year, the European Union announced plans to regulate AI technology within its borders. As the bloc has already highlighted with GDPR, its extensive legislation protecting online rights and privacy, the EU sees regulation of digital technology as one of the key challenges of current times.

The EU’s proposed regulation focuses particularly on so-called “high-risk” use cases of AI. The most prominent examples, such as facial recognition software, involve acute effects on individuals from bias or abuse. But AI applications can be risky on a broader scale, as well, particularly if they affect our options to address climate change.

To highlight this connection, we at Climate Change AI published a public comment on the proposed EU AI Act. We hope the final regulation will better reflect the many-faceted interplay between the potential of AI and the global challenge of a changing climate.

In particular, we believe the AI Act provides a unique opportunity to shape this interplay two ways:

First, climate change mitigation and adaptation should join other factors in informing whether an AI system is classified as “high-risk”. For example, if an AI solution leads to a substantial increase in greenhouse gas emissions, this should be reflected in its risk assessment.

Second, reporting requirements for high-risk AI systems should be expanded to include effects on greenhouse gas emissions. Specifically, solution providers should be required not just to report on emissions from powering the system’s computations, but also to describe projected emissions impacts—both positive and negative—of the applications the system enables. This approach would leverage the opportunity of reporting requirements for high-risk AI systems to collect much-needed data for decision-making on decarbonization strategies.

You can read our full commentary, with more details on these suggestions and others, on the website of the European Commission.

Introducing the CCAI blog

2021-10-28T00:00:00+00:00

CCAI’s core goals are all about serving our growing community of researchers and stakeholders who work at the intersection of climate change and AI. To that end, much of what we do is curate and produce content: we publish a newsletter; we run research workshops and host webinars; we post relevant papers, jobs, and news on Twitter and in our Circle community; we publish guides to working in this space; and so forth.

Even with all these channels, though, we’ve noticed a major gap: every so often, we’ve wanted to express an idea too big for a tweet or newsletter entry, but too informal for a highly produced web resource. Unfortunately, CCAI doesn’t have an appropriate outlet for this sort of content—or rather, it didn’t, until now.

We’re excited to announce the CCAI blog, which we hope will fill that gap. We plan for the blog to be a space for perspective pieces; for overviews of particularly interesting research; for guest appearances by the field’s many great thinkers; for roundups of links we’ve found enlightening; and for any other ideas, conversations, or reflections that we think our community will find thought-provoking and informative.

The blog won’t take over from either the newsletter or Circle. The newsletter will still serve as a periodic compilation of curated “quick hits,” and Circle will still be the place for community discussions and for sharing all manner of links, announcements, questions, and tips. The blog will cover new kinds of content, including expanded discussions of ideas and resources mentioned elsewhere in briefer form.

This project is still in its early days, and we expect it will keep evolving, so please do share any thoughts on what you’d like to see in this space.

In the meantime, we already have some great pieces in the pipeline. Expect more posts coming soon!

Climate Change AI

The Virtual Climate Change AI Summer School: 2024 Recap

The In-Person Climate Change AI Summer School: 2024 Recap

A Diverse Cohort

Training Sessions and Social Events

Project Development

Pitch Event

Goodbye, For Now

Project Descriptions:

Helping global grids work smarter, not harder

Smart grids are useful…and expensive

The GridWatch Dataset in Accra, Ghana

A call to help global grids work smarter, not harder

AI for Public and Planetary Health

Interview with Dr. Sara Khalid

Additional reading and resources

Unlocking Energy Flexibility of Residential Buildings

Notes

Announcing the 2023 CCAI Innovation Grants Awardees

NeurIPS 2023 Workshop: Blending new and existing knowledge systems

Overall Best Paper: Large Scale Masked Autoencoding for Reducing Label Requirements on SAR Data

Best ML Innovation: Real-time Carbon Footprint Minimization in Sustainable Data Centers with Reinforcement Learning

Best Pathway to Impact: Discovering Effective Policies for Land-Use Planning

Tips for great workshop papers

The Climate Change AI Summer School: 2023 Recap and 2024 Announcement

Virtual Summer School 2023

In-Person Summer School 2023

Summer School 2024

Teams and Team Members

Forecasting Land Degradation in Colombia

Precipitation Intelligence

Flooding Assessment for Coastal Environments (FACE)

Cassava Yield Optimization in Nigeria

WisePower Living

High Impact Climate and Weather Events Detection

Sustainability Report Chatbot (Susi)

Mentors

Using Machine Learning to Forecast the Weather and Climate

Intro

Forecasting the El Niño Southern Oscillation (ENSO) with Machine Learning

FourCastNet: A practical introduction to a state-of-the-art deep learning global weather emulator from Pathak et al., 2022

ClimateLearn: Machine Learning for Predicting Weather and Climate Extremes from Nguyen et al. 2023

Conclusions and Summary

Using Machine Learning to Increase Durability and Reduce Returns for Sports and Fashion Goods

Product Reviews to Measure Product Life Cycle and Quality

How Aspect-Based Sentiment Analysis Can Improve Product Design and Consumption Decisions

Fit:

Design:

Material:

Comfort:

Color:

Durability:

Our Results, Limitations, and Training Resources

Our mission

Introducing The ForestBench Project

Methodological Innovation: A Dual Approach to Data Collection

Community Engagement: Beyond Data Collection

Exploring the Conservation and Restoration of Blue Carbon Habitats: A Dive into Mangrove Forests

Adventures in Paraguay: A Journey of Collaboration and Conservation

A Warm Welcome from the Ministry

Into the Wild: The Defensores del Chaco National Park

A Lifelong Guardian: Don Silvino Gonzalez

Hope for Progress

Brazil: Guardians of the Amazon - The Kayapo Indigenous People

Bhutan: Planting a Million Dreams

Future Endeavors and Impact

Deep learning of nanoporous materials for chemical separations

Mapping Species From Crowdsourced Data Using Machine Learning

Using Machine Learning to Integrate Mangrove Restoration with Sustainable Aquaculture Intensification

Using Reinforcement Learning to Improve Energy Management for Grid-Interactive Buildings

Training off-line from fixed logs of an external policy

CCAI Core Team Profile: Priya Donti

Using AI-driven Yield Estimation to Improve Resilience of Malian Cotton Farmers

Climate Change Risks to Agri-Business Sector

The Dynamic Yield Estimation Map

The Team

Conclusion

Circularity in Fashion, powered by AI

What can AI do to accelerate a truly circular fashion?

AI can improve Design and Raw Materials Selection