Generative Modeling

Blog Posts

Workshop Papers

Venue Title
ICLR 2024 AI-driven emulation of ocean dynamics on sub-seasonal scales (Papers Track)
Abstract and authors: (click to expand)

Abstract: Climate forecasting systems rely on coupling atmospheric models to ocean and sea ice models. However, while there have recently been significant efforts to accelerate atmospheric models using AI, there have been very scarce efforts to accelerate the latter. As a result, climate forecasting systems still rely on expensive numerical simulations, which renders large-scale ensembling and probabilistic prediction cumbersome. To address this issue, we propose a large-scale AI model of ocean dynamics. Our method relies on a spherical neural operator to accurately capture the functional nature of ocean dynamics on the sphere. We empirically demonstrate that our model can accurately predict ocean dynamics for sub-seasonal horizons and outperforms the existing method. It offers a 60x speedup over the fastest numerical solver currently used for the task.

Authors: Suyash Bire (Massachusetts Institute of Technology); Jean Kossaifi (NVIDIA); Simone Silvestri (Massachusetts Institute of Technology); Nikola Kovachki (Nvidia Corp.); Kamyar Azizzadenesheli (Nvidia Corp.); Chris N Hill (MIT); Animashree Anandkumar (Caltech)

ICLR 2024 Extreme Precipitation Nowcasting using Transformer-based generative models (Papers Track)
Abstract and authors: (click to expand)

Abstract: This paper presents an innovative approach to extreme precipitation nowcasting by employing Transformer-based generative models, specifically VideoGPT with Extreme Value Loss (EVL) regularization. Leveraging a comprehensive dataset from the Royal Netherlands Meteorological Institute (KNMI), our study focuses on predicting short-term precipitation with high accuracy. We introduce a novel method for computing EVL without assuming fixed extreme representations, addressing the limitations of current models in capturing extreme weather events. We present both qualitative and quantitative analyses, demonstrating the superior performance of the proposed VideoGPT-EVL in generating accurate precipitation forecasts, especially when dealing with extreme precipitation events.

Authors: Cristian Meo (TUDelft); Mircea T Lica (Delft University of Technology); Ankush Roy (TUDelft); Zeina Boucher (TUDelft); Junzhe Yin (TUDelft); Yanbo Wang (Delft University of Technology); Ruben Imhoff (Deltares); Remko Uijlenhoet (TUDelft); Justin Dauwels (TU Delft)

ICLR 2024 Advancing Earth System Model Calibration: A Diffusion-Based Method (Papers Track) Honorable Mention
Abstract and authors: (click to expand)

Abstract: Understanding of climate impact on ecosystems globally requires site-specific model calibration. Here we introduce a novel diffusion-based uncertainty quantification (DBUQ) method for efficient model calibration. DBUQ is a score-based diffusion model that leverages Monte Carlo simulation to estimate the score function and evaluates a simple neural network to quickly generate samples for approximating parameter posterior distributions. DBUQ is stable, efficient, and can effectively calibrate the model given diverse observations, thereby enabling rapid and site-specific model calibration on a global scale. This capability significantly advances Earth system modeling and our understanding of climate impacts on Earth systems. We demonstrate DBUQ's capability in E3SM land model calibration at the Missouri Ozark AmeriFlux forest site. Both synthetic and real-data applications indicate that DBUQ produces accurate parameter posterior distributions similar to those generated by Markov Chain Monte Carlo sampling but with 30X less computing time. This efficiency marks a significant stride in model calibration, paving the way for more effective and timely climate impact analyses.

Authors: Yanfang Liu (Oak Ridge National Laboratory); Dan Lu (Oak Ridge National Laboratory); Zezhong Zhang (Oak Ridge National Laboratory); Feng Bao (Florida State University); Guannan Zhang (Oak Ridge National Laboratory)

ICLR 2024 Diffusion-Based Joint Temperature and Precipitation Emulation of Earth System Models (Papers Track)
Abstract and authors: (click to expand)

Abstract: Earth system models (ESMs) are the principal tools used in climate science to generate future climate projections under various atmospheric emissions scenarios on a global or regional scale. Generative deep learning approaches are suitable for emulating these tools due to their computational efficiency and ability, once trained, to generate realizations in a fraction of the time required by ESMs. We extend previous work that used a generative probabilistic diffusion model to emulate ESMs by targeting the joint emulation of multiple variables, temperature and precipitation, by a single diffusion model. Joint generation of multiple variables is critical to generate realistic samples of phenomena resulting from the interplay of multiple variables. The diffusion model emulator takes in the monthly mean-map of temperature and precipitation and produces the daily values of each of these variables that exhibit statistical properties similar to those generated by ESMs. Our results show the outputs from our extended model closely resemble those from ESMs on various climate metrics including dry spells and hot streaks, and that the joint distribution of temperature and precipitation in our sample closely matches those of ESMs.

Authors: Katie Christensen (Western Washington University); Lyric Otto (Western Washington University); Seth Bassetti (Utah State University); Claudia Tebaldi (Joint Global Change Research Institute); Brian Hutchinson (Western Washington University)

ICLR 2024 Fast non-stationary geospatial modelling with multiresolution (wavelet) Gaussian processes (Papers Track)
Abstract and authors: (click to expand)

Abstract: Climate modelling tasks involve assimilating large amounts of geospatial data from different sources, such as simulators and measurements from weather stations and satellites. These sources of data are weighted according to their uncertainty, so good quality uncertainty estimates are essential. Gaussian processes (GPs) offer flexible models with uncertainty estimates, and have a long track record of use in geospatial modelling. However, much of the research effort, including recent work on scalability, is focused on statistically stationary models, which are not suitable for many climatic variables, such as precipitation. Here we propose a novel, scalable, nonstationary GP model based upon discrete wavelets, and evaluate them on toy and real world data.

Authors: Talay M Cheema (University of Cambridge); Carl Edward Rasmussen (Cambridge University)

Abstract and authors: (click to expand)

Abstract: Access to smart meter data is essential to rapid and successful transitions to elec- trified grids, underpinned by flexibility delivered by low carbon technologies, such as electric vehicles (EV) and heat pumps, and powered by renewable energy. Yet little of this data is available for research and modelling purposes due consumer privacy protections. Whilst many are calling for raw datasets to be unlocked through regulatory changes, we believe this approach will take too long. Synthetic data addresses these challenges directly by overcoming privacy issues. In this pa- per, we present Faraday, a Variational Auto-encoder (VAE)-based model trained over 300 million smart meter data readings from an energy supplier in the UK, with information such as property type and low carbon technologies (LCTs) own- ership. The model produces household-level synthetic load profiles conditioned on these labels, and we compare its outputs against actual substation readings to show how the model can be used for real-world applications by grid modellers interested in modelling energy grids of the future.

Authors: Sheng Chai (Centre for Net Zero); Gus Chadney (Centre for Net Zero)

ICLR 2024 Forecasting Tropical Cyclones with Cascaded Diffusion Models (Papers Track)
Abstract and authors: (click to expand)

Abstract: As tropical cyclones become more intense due to climate change, the rise of Al-based modelling provides a more affordable and accessible approach compared to traditional methods based on mathematical models. This work leverages generative diffusion models to forecast cyclone trajectories and precipitation patterns by integrating satellite imaging, remote sensing, and atmospheric data. It employs a cascaded approach that incorporates three main tasks: forecasting, super-resolution, and precipitation modelling. The training dataset includes 51 cyclones from six major tropical cyclone basins from January 2019 - March 2023. Experiments demonstrate that the final forecasts from the cascaded models show accurate predictions up to a 36-hour rollout, with excellent Structural Similarity (SSIM) and Peak-Signal-To-Noise Ratio (PSNR) values exceeding 0.5 and 20 dB, respectively, for all three tasks. The 36-hour forecasts can be produced in as little as 30 mins on a single Nvidia A30/RTX 2080 Ti. This work also highlights the promising efficiency of Al methods such as diffusion models for high-performance needs in weather forecasting, such as tropical cyclone forecasting, while remaining computationally affordable, making them ideal for highly vulnerable regions with critical forecasting needs and financial limitations. Code accessible at \url{}.

Authors: Pritthijit Nath (Imperial College London); Pancham Shukla (Imperial College London); Shuai Wang (University of Delaware); Cesar Quilodran-Casas (Imperial College London)

ICLR 2024 DiffObs: Generative Diffusion for Global Forecasting of Satellite Observations (Papers Track)
Abstract and authors: (click to expand)

Abstract: This work presents an autoregressive generative diffusion model (DiffObs) to predict the global evolution of daily precipitation, trained on a satellite observational product, and assessed with domain-specific diagnostics. The model is trained to probabilistically forecast day-ahead precipitation. Nonetheless, it is stable for multi-month rollouts, which reveal a qualitatively realistic superposition of convectively coupled wave modes in the tropics. Cross-spectral analysis confirms successful generation of low frequency variations associated with the Madden--Julian oscillation, which regulates most subseasonal to seasonal predictability in the observed atmosphere, and convectively coupled moist Kelvin waves with approximately correct dispersion relationships. Despite secondary issues and biases, the results affirm the potential for a next generation of global diffusion models trained on increasingly sparse, and increasingly direct and differentiated observations of the world, for practical applications in subseasonal and climate prediction.

Authors: Jason Stock (Colorado State University); Jaideep Pathak (NVIDIA Corporation); Yair Cohen (NVIDIA Corporation); Mike Pritchard (NVIDIA Corporation); Piyush Garg (NVIDIA); Dale Durran (NVIDIA Corporation); Morteza Mardani (NVIDIA Corporation); Noah D Brenowitz (NVIDIA)

ICLR 2024 A Benchmark Dataset for Meteorological Downscaling (Proposals Track)
Abstract and authors: (click to expand)

Abstract: High spatial resolution in atmospheric representations is crucial across Earth science domains, but global reanalysis datasets like ERA5 often lack the detail to capture local phenomena due to their coarse resolution. Recent efforts have leveraged deep neural networks from computer vision to enhance the spatial resolution of meteorological data, showing promise for statistical downscaling. However, methodological diversity and insufficient comparisons with traditional downscaling techniques challenge these advancements. Our study introduces a benchmark dataset for statistical downscaling, utilizing ERA5 and the finer-resolution COSMO-REA6, to facilitate direct comparisons of downscaling methods for 2m temperature, global (solar) irradiance and 100m wind fields. Accompanying U-Net, GAN, and transformer models with a suite of evaluation metrics aim to standardize assessments and promote transparency and confidence in applying deep learning to meteorological downscaling.

Authors: Michael Langguth (Juelich Supercomputing Centre - Forschungszentrum Juelich); Paula Harder (Mila); Irene Schicker (Geos); Ankit Patnala (Juelich Supercomputing Centre - Forschungszentrum Juelich); Sebastian Lehner (GeoSphere Austria); Konrad Mayer (GeoSphere Austria); Markus Dabernig (GeoSphere Austria)

NeurIPS 2023 Combining deep generative models with extreme value theory for synthetic hazard simulation: a multivariate and spatially coherent approach (Papers Track)
Abstract and authors: (click to expand)

Abstract: Climate hazards can cause major disasters when they occur simultaneously as compound hazards. To understand the distribution of climate risk and inform adaptation policies, scientists need to simulate a large number of physically realistic and spatially coherent events. Current methods are limited by computational constraints and the probabilistic spatial distribution of compound events is not given sufficient attention. The bottleneck in current approaches lies in modelling the dependence structure between variables, as inference on parametric models suffers from the curse of dimensionality. Generative adversarial networks (GANs) are well-suited to such a problem due to their ability to implicitly learn the distribution of data in high-dimensional settings. We employ a GAN to model the dependence structure for daily maximum wind speed, significant wave height, and total precipitation over the Bay of Bengal, combining this with traditional extreme value theory for controlled extrapolation of the tails. Once trained, the model can be used to efficiently generate thousands of realistic compound hazard events, which can inform climate risk assessments for climate adaptation and disaster preparedness. The method developed is flexible and transferable to other multivariate and spatial climate datasets.

Authors: Alison M Peard (University of Oxford); Jim Hall (University of Oxford)

NeurIPS 2023 Climate Variable Downscaling with Conditional Normalizing Flows (Papers Track)
Abstract and authors: (click to expand)

Abstract: Predictions of global climate models typically operate on coarse spatial scales due to the large computational costs of climate simulations. This has led to a considerable interest in methods for statistical downscaling, a similar process to super-resolution in the computer vision context, to provide more local and regional climate information. In this work, we apply conditional normalizing flows to the task of climate variable downscaling. This approach allows for a probabilistic interpretation of the predictions, while also capturing the stochasticity inherent in the relationships among fine and coarse spatial scales. We showcase its successful performance on an ERA5 water content dataset for different upsampling factors. Additionally, we show that the method allows us to assess the predictive uncertainty in terms of standard deviation from the fitted conditional distribution mean.

Authors: Christina Elisabeth Winkler (Mila); Paula Harder (Mila); David Rolnick (McGill University, Mila)

NeurIPS 2023 RMM-VAE: a machine learning method for identifying probabilistic weather regimes targeted to a local-scale impact variable (Papers Track)
Abstract and authors: (click to expand)

Abstract: Identifying large-scale atmospheric patterns that modulate extremes in local-scale variables such as precipitation has the potential to improve long-term climate projections as well as extended-range forecasting skill. This paper proposes a novel probabilistic machine learning method, RMM-VAE, based on a variational autoencoder architecture for identifying weather regimes targeted to a local-scale impact variable. The new method is compared to three existing methods in the task of identifying robust weather regimes that are predictive of precipitation over Morocco while capturing the full phase space of atmospheric dynamics over the Mediterranean. RMM-VAE performs well across these different objectives, outperforming linear methods in reconstructing the full phase space and predicting the target variable, highlighting the potential benefit of applying the method to various climate applications such as downscaling and extended-range forecasting.

Authors: Fiona R Spuler (University of Reading); Marlene Kretschmer (Universität Leipzig); Yevgeniya Kovalchuck (University College London); Magdalena Balmaseda (ECMWF); Ted Shepherd (University of Reading)

NeurIPS 2023 Simulating the Air Quality Impact of Prescribed Fires Using a Graph Neural Network-Based PM2.5 Emissions Forecasting System (Papers Track)
Abstract and authors: (click to expand)

Abstract: The increasing size and severity of wildfires across western North America have generated dangerous levels of PM2.5 pollution in recent years. In a warming climate, expanding the use of prescribed fires is widely considered to be the most robust fire mitigation strategy. However, reliably forecasting the potential air quality impact from these prescribed fires, a critical ingredient in determining the fires’ location and time, at hourly to daily time scales remains a challenging problem. This paper proposes a novel integration of prescribed fire simulation with a spatio-temporal graph neural network-based PM2.5 forecasting model. The experiments in this work focus on determining the optimal time for implementing prescribed fires in California as well as quantifying the potential air quality trade-offs involved in conducting more prescribed fires outside the fire season.

Authors: Kyleen Liao (Saratoga High School); Jatan Buch (Columbia University); Kara D. Lamb (Columbia University); Pierre Gentine (Columbia University)

NeurIPS 2023 Fusion of Physics-Based Wildfire Spread Models with Satellite Data using Generative Algorithms (Papers Track)
Abstract and authors: (click to expand)

Abstract: Climate change has driven increases in wildfire prevalence, prompting development of wildfire spread models. Advancements in the use of satellites to detect fire locations provides opportunity to enhance fire spread forecasts from numerical models via data assimilation. In this work, a method is developed to infer the history of a wildfire from satellite measurements using a conditional Wasserstein Generative Adversarial Network (cWGAN), providing the information necessary to initialize coupled atmosphere-wildfire models in a physics-informed approach based on measurements. The cWGAN, trained with solutions from WRF-SFIRE, produces samples of fire arrival times (fire history) from the conditional distribution of arrival times given satellite measurements, and allows for assessment of prediction uncertainty. The method is tested on four California wildfires and predictions are compared against measured fire perimeters and reported ignition times. An average Sorensen's coefficient of 0.81 for the fire perimeters and an average ignition time error of 32 minutes suggests that the method is highly accurate.

Authors: Bryan Shaddy (University of Southern California); Deep Ray (University of Maryland); Angel Farguell (San Jose State University); Valentina Calaza (University of Southern California); Jan Mandel (University of Colorado Denver); James Haley (Cooperative Institute for Research in the Atmosphere); Kyle Hilburn (Cooperative Institute for Research in the Atmosphere); Derek Mallia (University of Utah); Adam Kochanski (San Jose State University); Assad Oberai (University of Southern California)

NeurIPS 2023 A 3D super-resolution of wind fields via physics-informed pixel-wise self-attention generative adversarial network (Papers Track)
Abstract and authors: (click to expand)

Abstract: To mitigate global warming, greenhouse gas sources need to be resolved at a high spatial resolution and monitored in time to ensure the reduction and ultimately elimination of the pollution source. However, the complexity of computation in resolving high-resolution wind fields left the simulations impractical to test different time lengths and model configurations. This study presents a preliminary development of a physics-informed super-resolution (SR) generative adversarial network (GAN) that super-resolves the three-dimensional (3D) low-resolution wind fields by upscaling x9 times. We develop a pixel-wise self-attention (PWA) module that learns 3D weather dynamics via a self-attention computation followed by a 2D convolution. We also employ a loss term that regularizes the self-attention map during pretraining, capturing the vertical convection process from input wind data. The new PWA SR-GAN shows the high-fidelity super-resolved 3D wind data, learns a wind structure at the high-frequency domain, and reduces the computational cost of a high-resolution wind simulation by x 89.7 times.

Authors: Takuya Kurihana (University of Chicago); Levente Klein (IBM Research); Kyongmin Yeo (IBM Research); Daniela Szwarcman (IBM Research); Bruce G Elmegreen (IBM Research); Surya Karthik Mukkavilli (IBM Research, Zurich); Johannes Schmude (IBM)

NeurIPS 2023 Inference of CO2 flow patterns--a feasibility study (Papers Track)
Abstract and authors: (click to expand)

Abstract: As the global deployment of carbon capture and sequestration (CCS) technology intensifies in the fight against climate change, it becomes increasingly imperative to establish robust monitoring and detection mechanisms for potential underground CO2 leakage, particularly through pre-existing or induced faults in the storage reservoir's seals. While techniques such as history matching and time-lapse seismic monitoring of CO2 storage have been used successfully in tracking the evolution of CO2 plumes in the subsurface, these methods lack principled approaches to characterize uncertainties related to the CO2 plumes' behavior. Inclusion of systematic assessment of uncertainties is essential for risk mitigation for the following reasons: (i) CO2 plume-induced changes are small and seismic data is noisy; (ii) changes between regular and irregular (e.g., caused by leakage) flow patterns are small; and (iii) the reservoir properties that control the flow are strongly heterogeneous and typically only available as distributions. To arrive at a formulation capable of inferring flow patterns for regular and irregular flow from well and seismic data, the performance of conditional normalizing flow will be analyzed on a series of carefully designed numerical experiments. While the inferences presented are preliminary in the context of an early CO2 leakage detection system, the results do indicate that inferences with conditional normalizing flows can produce high-fidelity estimates for CO2 plumes with or without leakage. We are also confident that the inferred uncertainty is reasonable because it correlates well with the observed errors. This uncertainty stems from noise in the seismic data and from the lack of precise knowledge of the reservoir's fluid flow properties.

Authors: Abhinav Prakash Gahlot (Georgia Institute of Technology); Huseyin Tuna Erdinc (Georgia Institute of Technology); Rafael Orozco (Georgia Institute of Technology); Ziyi Yin (Georgia Institute of Technology); Felix Herrmann (Georgia Institute of Technology)

NeurIPS 2023 Generative Nowcasting of Marine Fog Visibility in the Grand Banks area and Sable Island in Canada (Papers Track)
Abstract and authors: (click to expand)

Abstract: This study presents the application of generative deep learning techniques to evaluate marine fog visibility nowcasting using the FATIMA (Fog and turbulence interactions in the marine atmosphere) campaign observations collected during July 2022 in the North Atlantic in the Grand Banks area and vicinity of Sable Island (SI), northeast of Canada. The measurements were collected using the Vaisala Forward Scatter Sensor model FD70 and Weather Transmitter model WXT50, and Gill R3A ultrasonic anemometer mounted on the Research Vessel Atlantic Condor. To perform nowcasting, the time series of fog visibility (Vis), wind speed, dew point depression, and relative humidity with respect to water were preprocessed to have lagged time step features. Generative nowcasting of Vis time series for lead times of 30 and 60 minutes were performed using conditional generative adversarial networks (cGAN) regression at visibility thresholds of Vis < 1 km and < 10 km. Extreme gradient boosting (XGBoost) was used as a baseline method for comparison against cGAN. At the 30 min lead time, Vis was best predicted with cGAN at Vis < 1 km (RMSE = 0.151 km) and with XGBoost at Vis < 10 km (RMSE = 2.821 km). At the 60 min lead time, Vis was best predicted with XGBoost at Vis < 1 km (RMSE = 0.167 km) and Vis < 10 km (RMSE = 3.508 km), but the cGAN RMSE was similar to XGBoost. Despite nowcasting Vis at 30 min being quite difficult, the ability of the cGAN model to track the variation in Vis at 1 km suggests that there is potential for generative analysis of marine fog visibility using observational meteorological parameters.

Authors: Eren Gultepe (Southern Illinois University Edwardsville); Sen Wang (University of Notre Dame); Byron Blomquist (NOAA); Harindra Fernando (University of Notre Dame); Patrick Kreidl (University of North Florida); David Delene (University of North Dakota); Ismail Gultepe (Ontario Tech University)

NeurIPS 2023 Sustainability AI copilot: Analyze & ideate at scale to enable positive impact (Papers Track)
Abstract and authors: (click to expand)

Abstract: With the advances in large scale Foundation Models, web scale access to sustainability data, planetary scale satellite data, the opportunity for larger section of the world population to create positive climate impact can be activated by empowering everyone to ideate via AI copilots. The challenge is: How to enable more people to think & take action on climate & Sustainable Development goals?. We develop AI co-pilots to engage broader community for enabling impact at scale by democratizing climate thinking & ideation tools. We demonstrated how ideating with SAI transforms any seed idea into a holistic one, given the relation between climate & social economic aspects. SAI employs Language Models to represent the voice of the often neglected vulnerable people to the brainstorming discussion for inclusive climate action. We demonstrated how SAI can even create another AI that learns geospatial insights and offers advice to prevent humanitarian disasters from climate change. In this work, we conceptualized, designed, implemented & demonstrated Sustainability AI copilot (SAI) & innovated 4 use cases:- SAI enables sustainability enthusiasts to convert early stage budding thoughts into a robust holistic idea by creatively employing a chain of Large Language Models to think with six-thinking hats ideation. SAI can enables non-experts to become geospatial analysts by generating code to analyze planetary scale satellite data. SAI also ideates in multi-modal latent space to explore climate friendly product designs. SAI also enables human right activists to create awareness about inclusion of vulnerable and persons with disability in the climate conversation. SAI even creates AI apps for persons with disability. We demonstrated working prototypes at the project website, . Thus, SAI co-pilot empowers everyone to come together to ideate to make progress on climate and related sustainable development goals.

Authors: Rajagopal A (Indian Institute of Technology); Nirmala V (Queen Marys); Immanuel Raja (Karunya University); Arun V (NIT)

ICLR 2023 Improving global high-resolution Earth system model simulations of precipitation with generative adversarial networks (Papers Track)
Abstract and authors: (click to expand)

Abstract: Precipitation extremes are expected to become stronger and more frequent in response to anthropogenic global warming. Accurately projecting the ecological and socioeconomic impacts is an urgent task. Impact models are developed and calibrated with observation-based data but rely on Earth system model (ESM) output for future scenarios. ESMs, however, exhibit significant biases in their output because they cannot fully resolve complex cross-scale interactions of processes that produce precipitation cannot. State-of-the-art bias correction methods only address errors in the simulated frequency distributions, locally at every individual grid cell. Improving unrealistic spatial patterns of the ESM output, which would require spatial context, has not been possible so far. Here we show that a post-processing method based on physically constrained generative adversarial networks (GANs) can correct biases of a state-of-the-art global ESM both in local frequency distributions and in the spatial patterns at once. While our method improves local frequency distributions similarly well to a gold-standard ESM bias-adjustment framework, it strongly outperforms existing methods in correcting spatial patterns. Our study highlights the importance of physical constraints in neural networks for out-of-sample predictions in the context of climate change.

Authors: Philipp Hess (Technical University of Munich)

ICLR 2023 Graph Neural Network Generated Metal-Organic Frameworks for Carbon Capture (Proposals Track)
Abstract and authors: (click to expand)

Abstract: The level of carbon dioxide (CO2) in our atmosphere is rapidly rising and is projected to double today‘s levels to reach 1,000 ppm by 2100 under certain scenarios, primarily driven by anthropogenic sources. Technology that can capture CO2 from anthropogenic sources, remove from atmosphere and sequester it at the gigaton scale by 2050 is required stop and reverse the impact of climate change. Metal-organic frameworks (MOFs) have been a promising technology in various applications including gas separation as well as CO2 capture from point-source flue gases or removal from the atmosphere. MOFs offer unmatched surface area through their highly porous crystalline structure and MOF technology has potential to become a leading adsorption-based CO2 separation technology providing high surface area, structure stability and chemical tunability. Due to its complex structure, MOF crystal structure (atoms and bonds) cannot be easily represented in tabular format for machine learning (ML) applications whereas graph neural networks (GNN) have already been explored in representation of simpler chemical molecules. In addition to difficulty in MOF data representation, an infinite number of combinations can be created for MOF crystals, which makes ML applications more suitable to alleviate dependency on subject matter experts (SME) than conventional computational methods. In this work, we propose training of GNNs in variational autoencoder (VAE) setting to create an end-to-end workflow for the generation of new MOF crystal structures directly from the data within the crystallographic information files (CIFs) and conditioned by additional CO2 performance values.

Authors: Zikri Bayraktar (Schlumberger Doll Research); Shahnawaz Molla (Schlumberger Doll Research); Sharath Mahavadi (Schlumberger Doll Research)

ICLR 2023 DiffESM: Conditional Emulation of Earth System Models with Diffusion Models (Papers Track)
Abstract and authors: (click to expand)

Abstract: Earth System Models (ESMs) are essential tools for understanding the impact of human actions on Earth's climate. One key application of these models is studying extreme weather events, such as heat waves or dry spells, which have significant socioeconomic and environmental consequences. However, the computational demands of running a sufficient number of simulations to analyze the risks are often prohibitive. In this paper we demonstrate that diffusion models -- a class of generative deep learning models -- can effectively emulate the spatio-temporal trends of ESMs under previously unseen climate scenarios, while only requiring a small fraction of the computational resources. We present a diffusion model that is conditioned on monthly averages of temperature or precipitation on a 96x96 global grid, and produces daily values that are both realistic and consistent with those averages. Our results show that the output from our diffusion model closely matches the spatio-temporal behavior of the ESM it emulates in terms of the frequency of phenomena such as heat waves, dry spells, or rainfall intensity.

Authors: Seth Bassetti (Western Washington University); Brian Hutchinson (Western Washington University); Claudia Tebaldi (Joint Global Change Research Institute); Ben Kravitz (Indiana University)

ICLR 2023 Fourier Neural Operators for Arbitrary Resolution Climate Data Downscaling (Papers Track)
Abstract and authors: (click to expand)

Abstract: Running climate simulations informs us of future climate change. However, it is computationally expensive to resolve complex climate processes numerically. As one way to speed up climate simulations, neural networks have been used to downscale climate variables from fast-running low-resolution simulations. So far, all neural network downscaling models can only downscale input samples with a pre-defined upsampling factor. In this work, we propose a Fourier neural operator downscaling model. It trains with data of a small upsampling factor and then can zero-shot downscale its input to arbitrary unseen high-resolutions. Evaluated on Navier-Stokes equation solution data and ERA5 water content data, our downscaling model demonstrates better performance than widely used convolutional and adversarial generative super-resolution models in both learned and zero-shot downscaling. Our model's performance is further boosted when a constraint layer is applied. In the end, we show that by combining our downscaling model with a low-resolution numerical PDE solver, the downscaled solution outperforms the solution of the state-of-the-art high-resolution data-driven solver. Our model can be used to cheaply and accurately generate arbitrarily high-resolution climate simulation data with fast-running low-resolution simulation as input.

Authors: Qidong Yang (New York University); Paula Harder (Fraunhofer ITWM); Venkatesh Ramesh (University of Montreal, Mila); Alex Hernandez-Garcia (Mila - Quebec AI Institute); Daniela Szwarcman (IBM Research); Prasanna Sattigeri (IBM Research); Campbell D Watson (IBM Reserch); David Rolnick (McGill University, Mila)

NeurIPS 2022 Machine learning emulation of a local-scale UK climate model (Papers Track)
Abstract and authors: (click to expand)

Abstract: Climate change is causing the intensification of rainfall extremes. Precipitation projections with high spatial resolution are important for society to prepare for these changes, e.g. to model flooding impacts. Physics-based simulations for creating such projections are very computationally expensive. This work demonstrates the effectiveness of diffusion models, a form of deep generative models, for generating much more cheaply realistic high resolution rainfall samples for the UK conditioned on data from a low resolution simulation. We show for the first time a machine learning model that is able to produce realistic high-resolution rainfall predictions based on a physical model that resolves atmospheric convection, a key process behind extreme rainfall. By adding self-learnt, location-specific information to low resolution relative vorticity, quantiles and time-mean of the samples match well their counterparts from the high-resolution simulation.

Authors: Henry Addison (University of Bristol); Elizabeth Kendon (Met Office Hadley Centre); Suman Ravuri (DeepMind); Laurence Aitchison (University of Bristol); Peter Watson (Bristol)

NeurIPS 2022 Convolutional Neural Processes for Inpainting Satellite Images: Application to Water Body Segmentation (Papers Track)
Abstract and authors: (click to expand)

Abstract: The widespread availability of satellite images has allowed researchers to monitor the impact of climate on socio-economic and environmental issues through examples like crop and water body classification to measure food scarcity and risk of flooding. However, a common issue of satellite images is missing values due to measurement defects, which render them unusable by existing methods without data imputation. To repair the data, inpainting methods can be employed, which are based on classical PDEs or interpolation methods. Recently, deep learning approaches have shown promise in this realm, however many of these methods do not explicitly take into account the inherent spatio-temporal structure of satellite images. In this work, we cast satellite image inpainting as a meta-learning problem, and implement Convolutional Neural Processes (ConvNPs) in which we frame each satellite image as its own task or 2D regression problem. We show that ConvNPs outperform classical methods and state-of-the-art deep learning inpainting models on a scanline problem for LANDSAT 7 satellite images, assessed on a variety of in- and out-of-distribution images. Our results successfully match the performance of clean images on a downstream water body segmentation task in Canada.

Authors: Alexander Pondaven (Imperial College London); Mart Bakler (Imperial College London); Donghu Guo (Imperial College London); Hamzah Hashim (Imperial College London); Martin G Ignatov (Imperial college London); Samir Bhatt (Imperial College London); Seth Flaxman (Oxford); Swapnil Mishra (Imperial College London); Elie Alhajjar (USMA); Harrison Zhu (Imperial College London)

NeurIPS 2022 Don't Waste Data: Transfer Learning to Leverage All Data for Machine-Learnt Climate Model Emulation (Papers Track)
Abstract and authors: (click to expand)

Abstract: How can we learn from all available data when training machine-learnt climate models, without incurring any extra cost at simulation time? Typically, the training data comprises coarse-grained high-resolution data. But only keeping this coarse-grained data means the rest of the high-resolution data is thrown out. We use a transfer learning approach, which can be applied to a range of machine learning models, to leverage all the high-resolution data. We use three chaotic systems to show it stabilises training, gives improved generalisation performance and results in better forecasting skill. Our code is at

Authors: Raghul Parthipan (University of Cambridge); Damon Wischik (Univeristy of Cambridge)

NeurIPS 2022 Reconstruction of Grid Measurements in the Presence of Adversarial Attacks (Papers Track)
Abstract and authors: (click to expand)

Abstract: In efforts to mitigate the adverse effects of climate change, policymakers have set ambitious goals to reduce the carbon footprint of all sectors - including the electric grid. To facilitate this, sustainable energy systems like renewable generation must { be} deployed at high numbers throughout the grid. As these are highly variable in nature, the grid must be closely monitored so that system operators will have elevated situational awareness and can execute timely actions to maintain stable grid operations. With the widespread deployment of sensors like phasor measurement units (PMUs), an abundance of data is available for conducting accurate state estimation. However, due to the cyber-physical nature of the power grid, measurement data can be perturbed in an adversarial manner to enforce incorrect decision-making. In this paper, we propose a novel reconstruction method that leverages on machine learning constructs like CGAN and gradient search to recover the original states when subjected to adversarial perturbations. Experimental studies conducted on the practical IEEE 118-bus benchmark power system show that the proposed method can reduce errors due to perturbation by large margins (i.e. up to 100%).

Authors: Amirmohammad Naeini (York University); Samer El Kababji (Western University); Pirathayini Srikantha (York University)

NeurIPS 2022 Identifying Compound Climate Drivers of Forest Mortality with β-VAE (Papers Track)
Abstract and authors: (click to expand)

Abstract: Climate change is expected to lead to higher rates of forest mortality. Forest mortality is a complex phenomenon driven by the interaction of multiple climatic variables at multiple temporal scales, further modulated by the current state of the forest (e.g. age, stem diameter, and leaf area index). Identifying the compound climate drivers of forest mortality would greatly improve understanding and projections of future forest mortality risk. Observation data are, however, limited in accuracy and sample size, particularly in regard to forest state variables and mortality events. In contrast, simulations with state-of-the-art forest models enable the exploration of novel machine learning techniques for associating forest mortality with driving climate conditions. Here we simulate 160,000 years of beech, pine and spruce forest dynamics with the forest model FORMIND. We then apply β-VAE to learn disentangled latent representations of weather conditions and identify those that are most likely to cause high forest mortality. The learned model successfully identifies three characteristic climate representations that can be interpreted as different compound drivers of forest mortality.

Authors: Mohit Anand (Helmholtz Centre for Environmental Research - UFZ); Lily-belle Sweet (Helmholtz Centre for Environmental Research - UFZ); Gustau Camps-Valls (Universitat de València); Jakob Zscheischler (Helmholtz Centre for Environmental Research - UFZ)

NeurIPS 2022 Topic correlation networks inferred from open-ended survey responses reveal signatures of ideology behind carbon tax opinion (Papers Track)
Abstract and authors: (click to expand)

Abstract: Ideology can often render policy design ineffective by overriding what, at face value, are rational incentives. A timely example is carbon pricing, whose public support is strongly influenced by ideology. As a system of ideas, ideology expresses itself in the way people explain themselves and the world. As an object of study, ideology is then amenable to a generative modelling approach within the text-as-data paradigm. Here, we analyze the structure of ideology underlying carbon tax opinion using topic models. An idea, termed a topic, is operationalized as the fixed set of proportions with which words are used when talking about it. We characterize ideology through the relational structure between topics. To access this latent structure, we use the highly expressive Structural Topic Model to infer topics and the weights with which individual opinions mix topics. We fit the model to a large dataset of open-ended survey responses of Canadians elaborating on their support of or opposition to the tax. We propose and evaluate statistical measures of ideology in our data, such as dimensionality and heterogeneity. Finally, we discuss the implications of the results for transition policy in particular, and of our approach to analyzing ideology for computational social science in general.

Authors: Maximilian Puelma Touzel (Mila)

NeurIPS 2022 Controllable Generation for Climate Modeling (Papers Track)
Abstract and authors: (click to expand)

Abstract: Recent years have seen increased interest in modeling future climate trends, especially from the point of view of accurately predicting, understanding and mitigating downstream impacts. For instance, current state-of-the-art process-based agriculture models rely on high-resolution climate data during the growing season for accurate estimation of crop yields. However, high-resolution climate data for future climates is unavailable and needs to be simulated, and that too for multiple possible climate scenarios, which becomes prohibitively expensive via traditional methods. Meanwhile, deep generative models leveraging the expressivity of neural networks have shown immense promise in modeling distributions in high dimensions. Here, we cast the problem of simulation of climate scenarios in a generative modeling framework. Specifically, we leverage the GAN (Generative Adversarial Network) framework for simulating synthetic climate scenarios. We condition the model by quantifying the degree of ``extremeness" of the observed sample, which allows us to sample from different parts of the distribution. We demonstrate the efficacy of the proposed method on the CHIRPS precipitation dataset.

Authors: Moulik Choraria (University of Illinois at Urbana-Champaign); Daniela Szwarcman (IBM Research); Bianca Zadrozny (IBM Research); Campbell D Watson (IBM Reserch); Lav Varshney (UIUC: ECE)

NeurIPS 2022 Generative Modeling of High-resolution Global Precipitation Forecasts (Papers Track)
Abstract and authors: (click to expand)

Abstract: Forecasting global precipitation patterns and, in particular, extreme precipitation events is of critical importance to preparing for and adapting to climate change. Making accurate high-resolution precipitation forecasts using traditional physical models remains a major challenge in operational weather forecasting as they incur substantial computational costs and struggle to achieve sufficient forecast skill. Recently, deep-learning-based models have shown great promise in closing the gap with numerical weather prediction (NWP) models in terms of precipitation forecast skill, opening up exciting new avenues for precipitation modeling. However, it is challenging for these deep learning models to fully resolve the fine-scale structures of precipitation phenomena and adequately characterize the extremes of the long-tailed precipitation distribution. In this work, we present several improvements to the architecture and training process of a current state-of-the art deep learning precipitation model (FourCastNet) using a novel generative adversarial network (GAN) to better capture fine scales and extremes. Our improvements achieve superior performance in capturing the extreme percentiles of global precipitation, while comparable to state-of-the-art NWP models in terms of forecast skill at 1--2 day lead times. Together, these improvements set a new state-of-the-art in global precipitation forecasting.

Authors: James Duncan (University of California, Berkeley); Peter Harrington (Lawrence Berkeley National Laboratory (Berkeley Lab)); Shashank Subramanian (Lawrence Berkeley National Laboratory)

NeurIPS 2022 Synthesis of Realistic Load Data: Adversarial Networks for Learning and Generating Residential Load Patterns (Papers Track)
Abstract and authors: (click to expand)

Abstract: Responsible energy consumption plays a key role in reducing carbon footprint and CO2 emissions to tackle climate change. A better understanding of the residential consumption behavior using smart meter data is at the heart of the mission, which can inform residential demand flexibility, appliance scheduling, and home energy management. However, access to high-quality residential load data is still limited due to the cost-intensive data collection process and privacy concerns of data shar- ing. In this paper, we develop a Generative Adversarial Network (GAN)-based method to model the complex and diverse residential load patterns and generate synthetic yet realistic load data. We adopt a generation-focused weight selection method to select model weights to address the mode collapse problem and generate diverse load patterns. We evaluate our method using real-world data and demon- strate that it outperforms three representative state-of-the-art benchmark models in better preserving the sequence level temporal dependencies and aggregated level distributions of load patterns.

Authors: Xinyu Liang (Monash University); Hao Wang (Monash University)

AAAI FSS 2022 Towards Generating Large Synthetic Phytoplankton Datasets for Efficient Monitoring of Harmful Algal Blooms
Abstract and authors: (click to expand)

Abstract: Climate change is increasing the frequency and severity of harmful algal blooms (HABs), which cause significant fish deaths in aquaculture farms. This contributes to ocean pollution and greenhouse gas (GHG) emissions since dead fish are either dumped into the ocean or taken to landfills, which in turn negatively impacts the climate. Currently, the standard method to enumerate harmful algae and other phytoplankton is to manually observe and count them under a microscope. This is a time-consuming, tedious and error-prone process, resulting in compromised management decisions by farmers. Hence, automating this process for quick and accurate HAB monitoring is extremely helpful. However, this requires large and diverse datasets of phytoplankton images, and such datasets are hard to produce quickly. In this work, we explore the feasibility of generating novel high-resolution photorealistic synthetic phytoplankton images, containing multiple species in the same image, given a small dataset of real images. To this end, we employ Generative Adversarial Networks (GANs) to generate synthetic images. We evaluate three different GAN architectures: ProjectedGAN, FastGAN, and StyleGANv2 using standard image quality metrics. We empirically show the generation of high-fidelity synthetic phytoplankton images using a training dataset of only 961 real images. Thus, this work demonstrates the ability of GANs to create large synthetic datasets of phytoplankton from small training datasets, accomplishing a key step towards sustainable systematic monitoring of harmful algal blooms.

Authors: Nitpreet Bamra (University of Waterloo), Vikram Voleti (Mila, University of Montreal), Alexander Wong (University of Waterloo) and Jason Deglint (University of Waterloo)

AAAI FSS 2022 Wildfire Forecasting with Satellite Images and Deep Generative Model
Abstract and authors: (click to expand)

Abstract: Wildfire prediction has been one of the most critical tasks that humanities want to thrive at. While it plays a vital role in protecting human life, it is also difficult because of its stochastic and chaotic properties. We tackled the problem by interpreting a series of wildfire images into a video and used it to anticipate how the fire would behave in the future. However, creating video prediction models that account for the inherent uncertainty of the future is challenging. The bulk of published attempts are based on stochastic image-autoregressive recurrent networks, which raise various performance and application difficulties such as computational cost and limited efficiency on massive datasets. Another possibility is to use entirely latent temporal models that combine frame synthesis with temporal dynamics. However, due to design and training issues, no such model for stochastic video prediction has yet been proposed in the literature. This paper addresses these issues by introducing a novel stochastic temporal model whose dynamics are driven in a latent space. It naturally predicts video dynamics by allowing our lighter, more interpretable latent model to beat previous state-of-the-art approaches on the GOES-16 dataset. Results are compared using various benchmarking models.

Authors: Thai-Nam Hoang (University of Wisconsin - Madison), Sang Truong (Stanford University) and Chris Schmidt (University of Wisconsin - Madison)

AAAI FSS 2022 Contrastive Learning for Climate Model Bias Correction and Super-Resolution
Abstract and authors: (click to expand)

Abstract: Climate models often require post-processing in order to make accurate estimates of local climate risk. The most common post-processing applied is bias-correction and spatial resolution enhancement. However, the statistical methods typically used for this not only are incapable of capturing multivariate spatial correlation information but are also reliant on rich observational data often not available outside of developed countries, limiting their potential. Here we propose an alternative approach to this challenge based on a combination of image super resolution (SR) and contrastive learning generative adversarial networks (GANs). We benchmark performance against NASA’s flagship post-processed CMIP6 climate model product, NEX-GDDP. We find that our model successfully reaches a spatial resolution double that of NASA’s product while also achieving comparable or improved levels of bias correction in both daily precipitation and temperature. The resulting higher fidelity simulations of present and forward-looking climate can enable more local, accurate models of hazards like flooding, drought, and heatwaves.

Authors: Tristan Ballard (Sust Global) and Gopal Erinjippurath (Sust Global)

NeurIPS 2021 Towards Representation Learning for Atmospheric Dynamics (Papers Track)
Abstract and authors: (click to expand)

Abstract: The prediction of future climate scenarios under anthropogenic forcing is critical to understand climate change and to assess the impact of potentially counter-acting technologies. Machine learning and hybrid techniques for this prediction rely on informative metrics that are sensitive to pertinent but often subtle influences. For atmospheric dynamics, a critical part of the climate system, no well established metric exists and visual inspection is currently still often used in practice. However, this ``eyeball metric'' cannot be used for machine learning where an algorithmic description is required. Motivated by the success of intermediate neural network activations as basis for learned metrics, e.g. in computer vision, we present a novel, self-supervised representation learning approach specifically designed for atmospheric dynamics. Our approach, called AtmoDist, trains a neural network on a simple, auxiliary task: predicting the temporal distance between elements of a randomly shuffled sequence of atmospheric fields (e.g. the components of the wind field from reanalysis or simulation). The task forces the network to learn important intrinsic aspects of the data as activations in its layers and from these hence a discriminative metric can be obtained. We demonstrate this by using AtmoDist to define a metric for GAN-based super resolution of vorticity and divergence. Our upscaled data matches both visually and in terms of its statistics a high resolution reference closely and it significantly outperform the state-of-the-art based on mean squared error. Since AtmoDist is unsupervised, only requires a temporal sequence of fields, and uses a simple auxiliary task, it has the potential to be of utility in a wide range of applications.

Authors: Sebastian Hoffmann (University of Magdeburg); Christian Lessig (Otto-von-Guericke-Universitat Magdeburg)

NeurIPS 2021 Towards debiasing climate simulations using unsuperviserd image-to-image translation networks (Papers Track)
Abstract and authors: (click to expand)

Abstract: Climate models form the basis of a vast portion of earth system research, and inform our climate policy. Due to the complex nature of our climate system, and the approximations which must necessarily be made in simulating it, these climate models may not perfectly match observations. For further research, these outputs must be bias corrected against observations, but current methods of debiasing do not take into account spatial correlations. We evaluate unsupervised image-to-image translation networks, specifically the UNIT model architecture, for their ability to produce more spatially realistic debiasing than the standard techniques used in the climate community.

Authors: James Fulton (University of Edinburgh); Ben Clarke (Oxford University)

NeurIPS 2021 Mapping Post-Climate Change Biogeographical Regions with Deep Latent Variable Models (Papers Track)
Abstract and authors: (click to expand)

Abstract: Forecasting future changes to biodiversity due to shifts in climate is challenging due to nonlinear interactions between species as recorded in their presence/absence data. This work proposes using variational autoencoders with environmental covariates to identify low-dimensional structure in species’ joint co-occurrence patterns and leveraging this simplified representation to provide multivariate predictions of their habitat extent under future climate scenarios. We pursue a latent space clustering approach to map biogeographical regions of frequently co-occurring species and apply this methodology to a dataset from northern Belgium, generating predictive maps illustrating how these regions may expand or contract with changing temperature under a future climate scenario.

Authors: Christopher Krapu (Oak Ridge National Lab - Oak Ridge, TN)

NeurIPS 2021 WiSoSuper: Benchmarking Super-Resolution Methods on Wind and Solar Data (Papers Track)
Abstract and authors: (click to expand)

Abstract: The transition to green energy grids depends on detailed wind and solar forecasts to optimize the siting and scheduling of renewable energy generation. Operational forecasts from numerical weather prediction models, however, only have a spatial resolution of 10 to 20-km, which leads to sub-optimal usage and development of renewable energy farms. Weather scientists have been developing super-resolution methods to increase the resolution, but often rely on simple interpolation techniques or computationally expensive differential equation-based models. Recently, machine learning-based models, specifically the physics-informed resolution-enhancing generative adversarial network (PhIREGAN), have outperformed traditional downscaling methods. We provide a thorough and extensible benchmark of leading deep learning-based super-resolution techniques, including the enhanced super-resolution generative adversarial network (ESRGAN) and an enhanced deep super-resolution (EDSR) network, on wind and solar data. We accompany the benchmark with a novel public, processed, and machine learning-ready dataset for benchmarking super-resolution methods on wind and solar data.

Authors: Rupa Kurinchi-Vendhan (Caltech); Björn Lütjens (MIT); Ritwik Gupta (University of California, Berkeley); Lucien D Werner (California Institute of Technology); Dava Newman (MIT); Steven Low (California Institute of Technology)

NeurIPS 2021 On the use of Deep Generative Models for "Perfect" Prognosis Climate Downscaling (Proposals Track)
Abstract and authors: (click to expand)

Abstract: Deep Learning has recently emerged as a "perfect" prognosis downscaling technique to compute high-resolution fields from large-scale coarse atmospheric data. Despite their promising results to reproduce the observed local variability, they are based on the estimation of independent distributions at each location, which leads to deficient spatial structures, especially when downscaling precipitation. This study proposes the use of generative models to improve the spatial consistency of the high-resolution fields, very demanded by some sectoral applications (e.g., hydrology) to tackle climate change.

Authors: Jose González-Abad (Institute of Physics of Cantabria); Jorge Baño-Medina (Institute of Physics of Cantabria); Ignacio Heredia (Institute of Physics of Cantabria)

ICML 2021 Estimation of Corporate Greenhouse Gas Emissions via Machine Learning (Papers Track)
Abstract and authors: (click to expand)

Abstract: As an important step to fulfill the Paris Agreement and achieve net-zero emissions by 2050, the European Commission adopted the most ambitious package of climate impact measures in April 2021 to improve the flow of capital towards sustainable activities. For these and other international measures to be successful, reliable data is key. The ability to see the carbon footprint of companies around the world will be critical for investors to comply with the measures and hit climate neutrality. However, with only a small portion of companies volunteering to disclose their greenhouse gas (GHG) emissions, it is nearly impossible for investors to align their investment strategies with the measures. By training a machine learning model on disclosed GHG emissions, we are able to estimate the emissions of other companies globally who do not disclose their emissions. In this paper, we show that our model provides accurate estimates of corporate GHG emissions to investors such that they are able to align their investments with the regulatory measures and achieve net-zero goals.

Authors: You Han (Bloomberg L.P.); Achintya Gopal (Bloomberg LP); Liwen Ouyang (Bloomberg L.P.); Aaron Key (Bloomberg LP)

ICML 2021 Improving Image-Based Characterization of Porous Media with Deep Generative Models (Papers Track)
Abstract and authors: (click to expand)

Abstract: Micro- and nanoscale imaging are important for characterizing subsurface formations for carbon sequestration, shale gas recovery, and hydrogen storage. Common imaging techniques, however, are often sample-destructive, expensive, require high levels of expertise, or only acquire planar data. The resulting image datasets therefore may not allow for a representative estimation of rock properties. In this work, we address these challenges in image-based characterization of porous media using deep generative models. We present a machine learning workflow for characterizing porous media from limited imaging data. We develop methods for 3D image volume translation and synthesis from 2D training data, apply this method to grayscale and multimodal image datasets of sandstones and shales, and simulate flow through the generated volumes. Results show that the proposed image reconstruction and generation approaches produce realistic pore-scale 3D representations of rock samples using only 2D training data. The models proposed here expand our capabilities for characterization of rock samples and enable new understanding of pore-scale storage and recovery processes.

Authors: Timothy Anderson (Stanford University); Kelly Guan (Stanford University); Bolivia Vega (Stanford University); Laura Froute (Stanford University); Anthony Kovscek (Stanford University)

ICML 2021 Probabilistic Short-Term Low-Voltage Load Forecasting using Bernstein-Polynomial Normalizing Flows (Papers Track)
Abstract and authors: (click to expand)

Abstract: The transition to a fully renewable energy grid requires better forecasting of demand at the low-voltage level. However, high fluctuations and increasing electrification cause huge forecast errors with traditional point estimates. Probabilistic load forecasts take future uncertainties into account and thus enables various applications in low-carbon energy systems. We propose an approach for flexible conditional density forecasting of short-term load based on Bernstein-Polynomial Normalizing Flows where a neural network controls the parameters of the flow. In an empirical study with 363 smart meter customers, our density predictions compare favorably against Gaussian and Gaussian mixture densities and also outperform a non-parametric approach based on the pinball loss for 24h-ahead load forecasting for two different neural network architectures.

Authors: Marcel Arpogaus (Konstanz University of Applied Sciences); Marcus Voß (Technische Universität Berlin (DAI-Labor)); Beate Sick (ZHAW and University of Zurich); Mark Nigge-Uricher (Bosch.IO GmbH); Oliver Dürr (Konstanz University of Applied Sciences)

ICML 2021 Controlling Weather Field Synthesis Using Variational Autoencoders (Papers Track)
Abstract and authors: (click to expand)

Abstract: One of the consequences of climate change is an observed increase in the frequency of extreme climate events. That poses a challenge for weather forecast and generation algorithms, which learn from historical data but should embed an often uncertain bias to create correct scenarios. This paper investigates how mapping climate data to a known distribution using variational autoencoders might help explore such biases and control the synthesis of weather fields towards more extreme climate scenarios. We experimented using a monsoon-affected precipitation dataset from southwest India, which should give a roughly stable pattern of rainy days and ease our investigation. We report compelling results showing that mapping complex weather data to a known distribution implements an efficient control for weather field synthesis towards more (or less) extreme scenarios.

Authors: Dario Augusto Borges Oliveira (IBM Research); Jorge Luis Guevara Diaz (IBM Research); Bianca Zadrozny (IBM Research); Campbell Watson (IBM Reserch)

ICML 2021 A comparative study of stochastic and deep generative models for multisite precipitation synthesis (Papers Track)
Abstract and authors: (click to expand)

Abstract: Future climate change scenarios are usually hypothesized using simulations from weather generators. However, there only a few works comparing and evaluating multisite weather generators, and there is no existing work contrasting promising deep learning approaches for weather generation against classical stochastic weather generators. This study shows preliminary results evaluating stochastic weather generators and deep generative models for multisite precipitation synthesis. Using a variety of metrics, we compare two open source weather generators: XWeathergen (an extension of the Weathergen library) and RGeneratePrec, and two deep generative models: GAN and VAE. Our preliminary results can serve as a guide for improving the design of deep learning architectures and algorithms for the multisite precipitation synthesis task.

Authors: Jorge Luis Guevara Diaz (IBM Research); Dario Augusto Borges Oliveira (IBM Research); Bianca Zadrozny (IBM Research); Campbell Watson (IBM Reserch)

ICML 2021 Toward efficient calibration of higher-resolution Earth System Models (Papers Track) Best Paper: Pathway to Impact
Abstract and authors: (click to expand)

Abstract: Projections of future climate change to support decision-making require high spatial resolution, but this is computationally prohibitive with modern Earth system models (ESMs). A major challenge is the calibration (parameter tuning) process, which requires running large numbers of simulations to identify the optimal parameter values. Here we train a convolutional neural network (CNN) on simulations from two lower-resolution (and thus much less expensive) versions of the same ESM, and a smaller number of higher-resolution simulations. Cross-validated results show that the CNN's skill exceeds that of a climatological baseline for most variables with as few as 5-10 examples of the higher-resolution ESM, and for all variables (including precipitation) with at least 20 examples. This proof-of-concept study offers the prospect of significantly more efficient calibration of ESMs, by reducing the required CPU time for calibration by 20-40 %.

Authors: Christopher Fletcher (University of Waterloo); William McNally (University of Waterloo); John Virgin (University of Waterloo)

ICML 2021 Sky Image Prediction Using Generative Adversarial Networks for Solar Forecasting (Papers Track)
Abstract and authors: (click to expand)

Abstract: Large-scale integration of solar photovoltaics (PV) is challenged by high variability in its power output, mainly due to local and short-term cloud events. To achieve accurate solar forecasting, it is paramount to accurately predict the movement of clouds. Here, we use generative adversarial networks (GANs) to predict future sky images based on past sky image sequences and show that our trained model can generate realistic future sky images and capture the dynamics of clouds in the context frames. The generated images are then evaluated for a downstream solar forecasting task; results show promising performance.

Authors: Yuhao Nie (Stanford University); Andea Scott (Stanford University); Eric Zelikman (Stanford University); Adam Brandt (Stanford University)

ICML 2021 EVGen: Adversarial Networks for Learning Electric Vehicle Charging Loads and Hidden Representations (Papers Track)
Abstract and authors: (click to expand)

Abstract: The nexus between transportation, the power grid, and consumer behavior is much more pronounced than ever before as the race to decarbonize intensifies. Electrification in the transportation sector has led to technology shifts and rapid deployment of electric vehicles (EVs). The potential increase in stochastic and spatially heterogeneous charging load presents a unique challenge that is not well studied, and will have significant impacts on grid operations, emissions, and system reliability if not managed effectively. Realistic scenario generators can help operators prepare, and machine learning can be leveraged to this end. In this work, we develop generative adversarial networks (GANs) to learn distributions of electric vehicle (EV) charging sessions and disentangled representations. We show that this model successfully parameterizes unlabeled temporal and power patterns and is able to generate synthetic data conditioned on these patterns. We benchmark the generation capability of this model with Gaussian Mixture Models (GMMs), and empirically show that our proposed model framework is better at capturing charging distributions and temporal dynamics.

Authors: Robert Buechler (Stanford University); Emmanuel O Balogun (Stanford University); Arun Majumdar (Stanford University); Ram Rajagopal (Stanford University)

ICML 2021 Enhancing Laboratory-scale Flow Imaging of Fractured Geological Media with Deep Learning Super Resolution (Proposals Track)
Abstract and authors: (click to expand)

Abstract: Injection into deep geological formations is a promising approach for the utilization, sequestration, and removal from the atmosphere of CO2 emissions. Laboratory experiments are essential to characterize how CO2 flows and reacts in various types of geological media. We reproduce such dynamic injection processes while imaging using Computed Tomography (CT) at sufficient temporal resolution to visualize changes in the flow field. The resolution of CT, however, is on the order of 100's of micrometers and insufficient to characterize fine-scale reaction-induced alterations to micro-fractures. Super resolution deep learning is, therefore, an essential tool to improve spatial resolution of dynamic CT images. We acquired and processed pairs of multi-scale low- and high-resolution CT rock images. We also show the performance of our baseline model on fractured rock images using peak signal to noise ratio and structural similarity index. Coupling dynamic CT imaging with deep learning results in visualization with enhanced spatial resolution of about a factor of 4 thereby enabling improved interpretation.

Authors: Manju Pharkavi Murugesu (Stanford University); Timothy Anderson (Stanford University); Niccolo Dal Santo (MathWorks, Inc.); Vignesh Krishnan (The MathWorks Ltd); Anthony Kovscek (Stanford University)

NeurIPS 2020 Formatting the Landscape: Spatial conditional GAN for varying population in satellite imagery (Papers Track)
Abstract and authors: (click to expand)

Abstract: Climate change is expected to reshuffle the settlement landscape: forcing people in affected areas to migrate, to change their lifeways, and continuing to affect demographic change throughout the world. Changes to the geographic distribution of population will have dramatic impacts on land use and land cover and thus constitute one of the major challenges of planning for climate change scenarios. In this paper, we explore a generative model framework for generating satellite imagery conditional on gridded population distributions. We make additions to the existing ALAE [30] architecture, creating a spatially conditional version: SCALAE. This method allows us to explicitly disentangle population from the model’s latent space and thus input custom population forecasts into the generated imagery. We postulate that such imagery could then be directly used for land cover and land use change estimation using existing frameworks, as well as for realistic visualisation of expected local change. We evaluate the model by comparing pixel and semantic reconstructions, as well as calculate the standard FID metric. The results suggest the model captures population distributions accurately and delivers a controllable method to generate realistic satellite imagery.

Authors: Tomas Langer (Intuition Machines); Natalia Fedorova (Max Planck Institute for Evolutionary Anthropology); Ron Hagensieker (

NeurIPS 2020 Predicting Landsat Reflectance with Deep Generative Fusion (Papers Track)
Abstract and authors: (click to expand)

Abstract: Public satellite missions are commonly bound to a trade-off between spatial and temporal resolution as no single sensor provides fine-grained acquisitions with frequent coverage. This hinders their potential to assist vegetation monitoring or humanitarian actions, which require detecting rapid and detailed terrestrial surface changes. In this work, we probe the potential of deep generative models to produce high-resolution optical imagery by fusing products with different spatial and temporal characteristics. We introduce a dataset of co-registered Moderate Resolution Imaging Spectroradiometer (MODIS) and Landsat surface reflectance time series and demonstrate the ability of our generative model to blend coarse daily reflectance information into low-paced finer acquisitions. We benchmark our proposed model against state-of-the-art reflectance fusion algorithms.

Authors: Shahine Bouabid (University of Oxford); Jevgenij Gamper (Cervest Ltd.)

NeurIPS 2020 NightVision: Generating Nighttime Satellite Imagery from Infra-Red Observations (Papers Track)
Abstract and authors: (click to expand)

Abstract: The recent explosion in applications of machine learning to satellite imagery often rely on visible images and therefore suffer from a lack of data during the night. The gap can be filled by employing available infra-red observations to generate visible images. This work presents how Deep Learning can be applied successfully to create those images by using U-Net based architectures. The proposed methods show promising results, achieving a structural similarity index (SSIM) up to 86\% on an independent test set and providing visually convincing output images, generated from infra-red observations.

Authors: Paula Harder (Fraunhofer ITWM); William Jones (University of Oxford); Redouane Lguensat (LSCE-IPSL); Shahine Bouabid (University of Oxford); James Fulton (University of Edinburgh); Dánnell Quesada-Chacón (Technische Universität Dresden); Aris Marcolongo (University of Bern); Sofija Stefanovic (University of Oxford); Yuhan Rao (North Carolina Institute for Climate Studies); Peter Manshausen (University of Oxford); Duncan Watson-Parris (University of Oxford)

NeurIPS 2020 EarthNet2021: A novel large-scale dataset and challenge for forecasting localized climate impacts (Papers Track)
Abstract and authors: (click to expand)

Abstract: Climate change is global, yet its concrete impacts can strongly vary between different locations in the same region. Seasonal weather forecasts currently operate at the mesoscale (> 1 km). For more targeted mitigation and adaptation, modelling impacts to < 100 m is needed. Yet, the relationship between driving variables and Earth’s surface at such local scales remains unresolved by current physical models. Large Earth observation datasets now enable us to create machine learning models capable of translating coarse weather information into high-resolution Earth surface forecasts encompassing localized climate impacts. Here, we define high-resolution Earth surface forecasting as video prediction of satellite imagery conditional on mesoscale weather forecasts. Video prediction has been tackled with deep learning models. Developing such models requires analysis-ready datasets. We introduce EarthNet2021, a new, curated dataset containing target spatio-temporal Sentinel 2 satellite imagery at 20 m resolution, matched with high-resolution topography and mesoscale (1.28 km) weather variables. With over 32000 samples it is suitable for training deep neural networks. Comparing multiple Earth surface forecasts is not trivial. Hence, we define the EarthNetScore, a novel ranking criterion for models forecasting Earth surface reflectance. For model intercomparison we frame EarthNet2021 as a challenge with four tracks based on different test sets. These allow evaluation of model validity and robustness as well as model applicability to extreme events and the complete annual vegetation cycle. In addition to forecasting directly observable weather impacts through satellite-derived vegetation indices, capable Earth surface models will enable downstream applications such as crop yield prediction, forest health assessments, coastline management, or biodiversity monitoring. Find data, code, and how to participate at .

Authors: Christian Requena-Mesa (Computer Vision Group, Friedrich Schiller University Jena; DLR Institute of Data Science, Jena; Max Planck Institute for Biogeochemistry, Jena); Vitus Benson (Max-Planck-Institute for Biogeochemistry); Jakob Runge (Institute of Data Science, German Aerospace Center (DLR)); Joachim Denzler (Computer Vision Group, Friedrich Schiller University Jena, Germany); Markus Reichstein (Max Planck Institute for Biogeochemistry, Jena; Michael Stifel Center Jena for Data-Driven and Simulation Science, Jena)

NeurIPS 2020 VConstruct: Filling Gaps in Chl-a Data Using a Variational Autoencoder (Papers Track)
Abstract and authors: (click to expand)

Abstract: Remote sensing of Chlorophyll-a is vital in monitoring climate change. Chlorphylla measurements give us an idea of the algae concentrations in the ocean, which lets us monitor ocean health. However, a common problem is that the satellites used to gather the data are commonly obstructed by clouds and other artifacts. This means that time series data from satellites can suffer from spatial data loss. There are a number of algorithms that are able to reconstruct the missing parts of these images to varying degrees of accuracy, with Data INterpolating Empirical Orthogonal Functions (DINEOF) being the current standard. However, DINEOF is slow, suffers from accuracy loss in temporally homogenous waters, reliant on temporal data, and only able to generate a single potential reconstruction. We propose a machine learning approach to reconstruction of Chlorophyll-a data using a Variational Autoencoder (VAE). Our accuracy results to date are competitive with but slightly less accurate than DINEOF. We show the benefits of our method including vastly decreased computation time and ability to generate multiple potential reconstructions. Lastly, we outline our planned improvements and future work.

Authors: Matthew Ehrler (University of Victoria); Neil Ernst (University of Victoria)

NeurIPS 2020 Loosely Conditioned Emulation of Global Climate Models With Generative Adversarial Networks (Papers Track)
Abstract and authors: (click to expand)

Abstract: Climate models encapsulate our best understanding of the Earth system, allowing research to be conducted on its future under alternative assumptions of how human-driven climate forces are going to evolve. An important application of climate models is to provide metrics of mean and extreme climate changes, particularly under these alternative future scenarios, as these quantities drive the impacts of climate on society and natural systems. Because of the need to explore a wide range of alternative scenarios and other sources of uncertainties in a computationally efficient manner, climate models can only take us so far, as they require significant computational resources, especially when attempting to characterize extreme events, which are rare and thus demand long and numerous simulations in order to accurately represent their changing statistics. Here we use deep learning in a proof of concept that lays the foundation for emulating global climate model output for different scenarios. We train two "loosely conditioned" Generative Adversarial Networks (GANs) that emulate daily precipitation output from a fully coupled Earth system model: one GAN modeling Fall-Winter behavior and the other Spring-Summer. Our GANs are trained to produce spatiotemporal samples: 32 days of precipitation over a 64x128 regular grid discretizing the globe. We evaluate the generator with a set of related performance metrics based upon KL divergence, and find the generated samples to be nearly as well matched to the test data as the validation data is to test. We also find the generated samples to accurately estimate the mean number of dry days and mean longest dry spell in the 32 day samples. Our trained GANs can rapidly generate numerous realizations at a vastly reduced computational expense, compared to large ensembles of climate models, which greatly aids in estimating the statistics of extreme events.

Authors: Alex Ayala (Western Washington University); Chris Drazic (Western Washington University); Brian Hutchinson (Western Washington University); Ben Kravitz (Indiana University); Claudia Tebaldi (Joint Global Change Research Institute)

NeurIPS 2020 A Generative Adversarial Gated Recurrent Network for Power Disaggregation & Consumption Awareness (Papers Track)
Abstract and authors: (click to expand)

Abstract: Separating the household aggregated power signal into its additive sub-components is called energy (power) disaggregation or Non-Intrusive Load Monitoring. NILM can play an instrumental role as a driver towards consumer energy consumption awareness and behavioral change. In this paper, we propose EnerGAN++, a model based on GANs for robust energy disaggregation. We propose a unified autoencoder (AE) and GAN architecture, in which the AE achieves a non-linear power signal source separation. The discriminator performs sequence classification, using a recurrent CNN to handle the temporal dynamics of an appliance energy consumption time series. Experimental results indicate the proposed method’s superiority compared to the current state of the art.

Authors: Maria Kaselimi (National Technical University of Athens); Athanasios Voulodimos (University of West Attica); Nikolaos Doulamis (National Technical University of Athens); Anastasios Doulamis (Technical University of Crete); Eftychios Protopapadakis (National Technical University of Athens)