Panel: Practical use-inspired climate technology - Dr. Peetak Mitra - Dr. Sherrie Wang - Sherif Elsayed-Ali - Anastasia Volkova

We introduce EarthPT -- an Earth Observation (EO) pretrained transformer. EarthPT is a 700 million parameter decoding transformer foundation model trained in an autoregressive self-supervised manner and developed specifically with EO use-cases in mind. We demonstrate that EarthPT is an effective forecaster that can accurately predict future pixel-level surface reflectances across the 400-2300 nm range well into the future. For example, forecasts of the evolution of the Normalised Difference Vegetation Index (NDVI) have a typical error of approximately 0.05 (over a natural range of -1 -> 1) at the pixel level over a five month test set horizon, out-performing simple phase-folded models based on historical averaging. We also demonstrate that embeddings learnt by EarthPT hold semantically meaningful information and could be exploited for downstream tasks such as highly granular, dynamic land use classification. Excitingly, we note that the abundance of EO data provides us with -- in theory -- quadrillions of training tokens. Therefore, if we assume that EarthPT follows neural scaling laws akin to those derived for Large Language Models (LLMs), there is currently no data-imposed limit to scaling EarthPT and other similar ‘Large Observation Models.’

Mapping aquaculture ponds is critical for restoration, conservation, and climate adaptation efforts. Aquaculture can contribute to high levels of water pollution from untreated effluent and negatively impact coastal ecosystems. Large-scale aquaculture is also a significant driver in mangrove deforestation, thus reducing the world’s carbon sinks and exacerbating the effects of climate change. However, finding and mapping these ponds on the ground can be highly labor and time-intensive. Most existing automated techniques are focused only on spatial location and do not consider production intensification, which is also crucial to understanding their impact on the surrounding ecosystem. We can classify them into two main types: a) Extensive ponds, which are large, irregularly-shaped ponds that rely on natural productivity, and b) intensive ponds which are smaller and regularly shaped. Intensive ponds use machinery such as aerators that maximize production and also result in the characteristic presence of air bubbles on the pond’s surface. The features of these two types of ponds make them distinguishable and detectable from satellite imagery.In this tutorial, we will discuss types of aquaculture ponds in detail and demonstrate how they can be detected and classified using satellite imagery. The tutorial will introduce an open dataset of human-labeled aquaculture ponds in the Philippines and Indonesia. Using this dataset, the tutorial will use semantic segmentation to map out similar ponds over an entire country and classify them as either extensive or intensive, going through the entire process of i) satellite imagery retrieval, ii) preprocessing these images into a training-ready dataset, iii) model training, and iv) finally model rollout on a sample area. Throughout, the tutorial will leverage PyTorch Lightning, a machine learning framework that provides a simplified and streamlined interface for model experimentation and deployment. This tutorial aims to discuss the relevance of aquaculture ponds in climate adaptation and equip users with the necessary inputs and tools to perform their own ML-powered earth observation projects.

Transportation contributes to 29% of all greenhouse gas (GHG) emissions in the US, of which 58% are from light-duty vehicles and 28% from medium-to-heavy duty vehicles (MHDVs) [1]. Battery electric vehicles (EVs) emit 90% less life cycle GHGs than their internal combustion engine (ICEV) counterparts [2], but currently only comprise 2% of all vehicles in the U.S. EVs thus represent a crucial step in decarbonizing road transportation. One major challenge in replacing ICEVs with EVs at scale is the ability to charge a large number of EVs within the constraints of power systems in a cost-effective way. This is an especially prominent problem for MHDVs used in commercial fleets such as shuttle buses and delivery trucks, as they generally require more energy to complete assigned trips compared to light-duty vehicles. In this tutorial, we describe the myriad challenges in charging EVs at scale and define common objectives such as minimizing total load on power systems, minimizing fleet operating costs, as well as maximizing vehicle state of charge and onsite photovoltaic energy usage. We discuss common constraints such as vehicle trip energy requirements, charging station power limits, and limits on vehicles’ time to charge between trips. We survey several different methods to formulate EV charging and energy dispatch as a mathematically solvable optimization problem, using tools such as convex optimization, Markov decision process (MDP), and reinforcement learning (RL). We introduce a commercial application of model-based predictive control (MPC) algorithm, ZEVi (Zero Emission Vehicle intelligence), which solves optimal energy dispatch strategies for charging sessions of commercial EV fleets. Using a synthetic dataset modeled after a real fleet of electric school buses, we engage the audience with a hands-on exercise applying ZEVi to find the optimal charging strategy for a commercial fleet. Lastly, we briefly discuss other contexts in which methods originating from process control and deep learning, like MPC and RL, can be applied to solve problems related to climate change mitigation and adaptation. With the examples provided in this tutorial, we hope to inspire the audience to come up with their own creative ways to apply these methods in different fields within the climate domain. References[1] EPA (2023). Inventory of U.S. Greenhouse Gas Emissions and Sinks: 1990-2021. U.S. Environmental Protection Agency, EPA 430-R-23-002. [2] Verma, S., Dwivedi, G., & Verma, P. (2022). Life cycle assessment of electric vehicles in comparison to combustion engine vehicles: A review. Materials Today: Proceedings, 49, 217-222.

Machine learning models for parsing remote sensing data have a wide range of societally relevant applications, but labels used to train these models can be difficult or impossible to acquire. This challenge has spurred research into self-supervised learning for remote sensing data. Current self-supervised learning approaches for remote sensing data draw significant inspiration from techniques applied to natural images. However, remote sensing data has important differences from natural images -- for example, the temporal dimension is critical for many tasks and data is collected from many complementary sensors. We show we can create significantly smaller performant models by designing architectures and self-supervised training techniques specifically for remote sensing data. We introduce the Pretrained Remote Sensing Transformer (Presto), a transformer-based model pre-trained on remote sensing pixel-timeseries data. Presto excels at a wide variety of globally distributed remote sensing tasks and performs competitively with much larger models while requiring far less compute. Presto can be used for transfer learning or as a feature extractor for simple models, enabling efficient deployment at scale.

We propose a framework for applying reinforcement learning to contextual two-stage stochastic optimization and apply this framework to the problem of energy market bidding of an off-shore wind farm. Reinforcement learning could potentially be used to learn close to optimal solutions for first stage variables of a two-stage stochastic program under different contexts. Under the proposed framework, these solutions would be learned without having to solve the full two-stage stochastic program. We present initial results of training using the DDPG algorithm and present intended future steps to improve performance.

Satellite-based remote sensing is instrumental in the monitoring and mitigation of the effects of anthropogenic climate change. Large scale, high resolution data derived from these sensors can be used to inform intervention and policy decision making, but the timeliness and accuracy of these interventions is limited by use of optical data, which cannot operate at night and is affected by adverse weather conditions. Synthetic Aperture Radar (SAR) offers a robust alternative to optical data, but its associated complexities limit the scope of labelled data generation for traditional deep learning. In this work, we apply a self-supervised pretraining scheme, masked autoencoding, to SAR amplitude data covering 8.7\% of the Earth's land surface area, and tune the pretrained weights on two downstream tasks crucial to monitoring climate change - vegetation cover prediction and land cover classification. We show that the use of this pretraining scheme reduces labelling requirements for the downstream tasks by more than an order of magnitude, and that this pretraining generalises geographically, with the performance gain increasing when tuned downstream on regions outside the pretraining set. Our findings significantly advance climate change mitigation by facilitating the development of task and region-specific SAR models, allowing local communities and organizations to deploy tailored solutions for rapid, accurate monitoring of climate change effects.

Fireside Chat: LLMs and their Implications for Climate Change - Dr. David Rolnick - Dr. Priya Donti - Dr. Sasha Luccioni

Keynote: Tanya Berger-Wolf

The growing penetration of intermittent, renewable generation in US power grids results in increased operational uncertainty. In that context, accurate forecasts are critical, especially for wind generation, which exhibits large variability and is historically harder to predict. To overcome this challenge, this work proposes a novel Bundle-Predict-Reconcile (BPR) framework that integrates asset bundling, machine learning, and forecast reconciliation techniques to accurately predict wind power at the asset, bundle, and fleet level. Notably, our approach effectively introduces an auxiliary learning task (predicting the bundle-level time series) to help the main learning tasks (fleet-level time series) and proposes new asset-bundling criteria to capture the spatio-temporal dynamics of wind power time series. Extensive numerical experiments are conducted on an industry-size dataset of wind farms, demonstrating the benefits of BPR, which consistently and significantly improves forecast accuracy over the baseline approach, especially at the fleet level.

Understanding why travel behavior differs between residents of urban centers and suburbs is key to sustainable urban planning. Especially in light of rapid urban growth, identifying housing locations that minimize travel demand and induced emissions is crucial to mitigate climate change. While the built environment plays an important role, the precise impact on travel behavior is obfuscated by residential self-selection.To address this issue, we propose a double machine learning approach to obtain unbiased, spatially-explicit estimates of the effect of the built environment on travel-related emissions for each neighborhood by controlling for residential self-selection. We examine how socio-demographics and travel-related attitudes moderate the effect and how it decomposes across the 5Ds of the built environment. Based on a case study for Berlin and the travel diaries of 32,000 residents, we find that the built environment causes household travel-related emissions to differ by a factor of almost two between central and suburban neighborhoods in Berlin. To highlight the practical importance for urban climate mitigation, we evaluate current plans for 64,000 new residential units in terms of total induced transport emissions. Our findings underscore the significance of compact development to decarbonize the transport sector.

How areas of land are allocated for different uses, such as forests, urban, and agriculture, has a large effect on carbon balance, and therefore climate change. Based on available historical data on changes in land use and a simulation of carbon emissions/absorption, a surrogate model can be learned that makes it possible to evaluate the different options available to decision-makers efficiently. An evolutionary search process can then be used to discover effective land-use policies for specific locations. Such a system was built on the Project Resilience platform and evaluated with the Land Use Harmonization dataset and the BLUE simulator. It generates Pareto fronts that trade off carbon impact and amount of change customized to different locations, thus providing a potentially useful tool for land-use planning.

The increasing size and severity of wildfires across western North America have generated dangerous concentrations 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 spatial-temporal graph neural network-based PM2.5 forecasting. The experiments in this work focus on determining the optimal time for implementing prescribed fires in California as well as on quantifying the potential trade-offs involved in conducting more prescribed fires outside the fire season.

Data-driven weather prediction (DDWP) models are increasingly becoming popular for weather forecasting. While DDWPs primarily forecast prognostic parameters, many diagnostic meteorological parameters (such as precipitation) are dependent on the most recent weather state and are modeled by learning a data-driven functional mapping of the current meteorological state (c.f. FourCastNet). However, the cost of training bespoke models for diagnostic variables can scale significantly and further limit the use during operationalizing these forecasts. This presents an opportunity to learn dense representations of essential meteorological parameters in a latent space, and using learned representations to model diagnostic parameters, or any other dependent variables. Using learned representations of weather allows for efficient prediction of dependent variables, while dramatically lowering the training cost for such models as well. In this paper, we present one such weather embedding model, WeatherX, trained on decades of reanalysis data that is used to train multiple diagnostic variables. The results indicate that models trained using learned representations of weather offer performance comparable to bespoke models, while leading to significant reduction in resource utilization during training and inference. Further lower memory footprint during operationalization leads to additional gain of running larger ensembles during inference thereby further improving uncertainty quantification of the said forecasts.

Panel: Local and indigenous knowledge systems - Chamisa Edmo - Kidane Woldemariyam - Alex Desiga - Tracy Monteith

Monitoring organic matter is pivotal for maintaining soil health and can help inform sustainable soil management practices. While sensor-based soil information offers higher-fidelity and reliable insights into organic matter changes, sampling and measuring sensor data is cost-prohibitive. We propose a multi-modal, scalable framework that can estimate organic matter from remote sensing data, a more readily available data source while leveraging sparse soil information for improving generalization. Using the sensor data, we preserve underlying causal relations among sensor attributes and organic matter. Simultaneously we leverage inherent structure in the data and train the model to discriminate among domains using contrastive learning. This causal and contrastive constraint minimization ensures improved generalization and adaptation to other domains. We also shed light on the interpretability of the framework by identifying attributes that are important for improving generalization. Identifying these key soil attributes that affect organic matter will aid in efforts to standardize data collection efforts.

As machine learning workloads significantly increase energy consumption, sustainable data centers with low carbon emissions are becoming a top priority for governments and corporations worldwide. There is a pressing need to optimize energy usage in these centers, especially considering factors like cooling, load flexibility based on renewable energy availability, and battery storage utilization. The challenge arises due to the interdependencies of these strategies with fluctuating external factors such as weather and grid carbon intensity. Although there's currently no real-time solution that addresses all these aspects, our proposed Data Center Carbon Footprint Reduction (DCCFR) framework, based on multi-agent Reinforcement Learning (MARL), targets carbon footprint reduction, energy conservation, and cost. Our findings reveal that DCCFR's MARL agents efficiently navigate these complexities, optimizing energy in real-time. Compared to the industry standard ASHRAE controller controlling HVAC for a year in various regions, DCCFR reduced carbon emissions, energy consumption, and energy costs by over 10% with EnergyPlus simulation.

Carbon capture technologies is an important tool for mitigating climate change. In recent years, polymer membrane separation methods have emerged as a promising technology for separating CO2 and other green house gases from the atmosphere. Designing new polymers for such tasks is quite difficult. In this work we look at machine learning based methods to search for new polymer designs optimized for CO2 separation. An ensemble ML models is trained on a large database of molecules to predict permeabilities of CO2/N2 and CO2/O2 pairs. We then use search based optimization to discover new polymers that surpass existing polymer designs. Simulations are then done to verify the predicted performance of the new designs. Overall result suggests that ML based search can be used to discover new polymers optimized for carbon capture.

High-resolution simulations such as the ICOsahedral Non-hydrostatic Large-Eddy Model (ICON-LEM) can be used to study the interactions between aerosols, clouds, and precipitation processes that currently represent the largest source of uncertainty involved in climate change projections. However, due to significant computational costs, it can only be employed for a limited period and area. While machine learning mitigates this, model uncertainties may affect reliability. To address this, we developed a neural network (NN) model powered with evidential learning to assess the data and model uncertainties. Our study focuses on estimating the rate at which small droplets (cloud droplets) collide and coalesce to become larger droplets (raindrops) – autoconversion rates -- as the key process in the precipitation formation, crucial to better understanding cloud responses to anthropogenic aerosols. The results show that the model performs reasonably well, with the inclusion of both aleatoric and epistemic uncertainty estimation, which improves the credibility of the model and provides useful insights for future improvement.

Existing ML-based atmospheric models are not suitable for climate prediction, which requires long-term stability and physical consistency. We present ACE (AI Climate Emulator), a 200M-parameter, autoregressive machine learning emulator of an existing comprehensive 100-km resolution global atmospheric model. The formulation of ACE allows evaluation of physical laws such as the conservation of mass and moisture. The emulator is stable for 10 years, nearly conserves column moisture without explicit constraints and faithfully reproduces the reference model's climate, outperforming a challenging baseline on over 80% of tracked variables. ACE requires nearly 100x less wall clock time and is 100x more energy efficient than the reference model using typically available resources.

Clouds containing ice particles play a crucial role in the climate system. Yet they remain a source of great uncertainty in climate models and future climate projections. In this work, we create a new observational constraint of regime-dependent ice microphysical properties at the spatio-temporal coverage of geostationary satellite instruments and the quality of active satellite retrievals. We achieve this by training a convolutional neural network on three years of SEVIRI and DARDAR data sets. This work will enable novel research to improve ice cloud process understanding and hence, reduce uncertainties in a changing climate and help assess geoengineering methods for cirrus clouds.

Greenhouse Gas (GHG) emissions datasets are often incomplete due to inconsistent reporting and poor transparency. Filling the gaps in these datasets allows for more accurate targeting of strategies to accelerate the reduction of GHG emissions. This study evaluates the potential of machine learning methods to automate the completion of GHG datasets. We use 3 datasets of increasing complexity with 18 different gap-filling methods and provide a guide to which methods are useful in which circumstances. If few dataset features are available, or the gap consists only of a missing time step in a record, then simple interpolation is often the most accurate method and complex models should be avoided. However, if more features are available and the gap involves non-reporting emitters, then machine learning methods can be more accurate than simple extrapolation. Furthermore, the secondary output of feature importance from complex models allows for data collection prioritisation to accelerate the improvement of datasets. Graph based methods are particularly scalable due to the ease of updating predictions given new data and incorporating multimodal data sources. This study can serve as a guide to the community upon which to base ever more integrated frameworks for automated detailed GHG emissions estimations, and implementation guidance is available at https://hackmd.io/@luke-scot/ML-for-GHG-database-completion.

Consumer products contribute to >75% of global greenhouse gas (GHG) emissions, primarily through indirect contributions from the supply chain. Measurement of GHG emissions associated with products is crucial to quantify the impact of GHG emission abatement actions. Life cycle assessment (LCA), the scientific discipline for measuring GHG emissions, estimates the environmental impact of a product. Scaling LCA to millions of products is challenging as it requires extensive manual analysis by domain experts. To avoid repetitive analysis, environmental impact factors (EIF) of common materials and products are published for use by experts. However, finding appropriate EIFs for even a single product can require hundreds of hours of manual work, especially for complex products. We present Flamingo, an algorithm that leverages neural language models to automatically identify an appropriate EIF given a text description. A key challenge in automation is that EIF databases are incomplete. Flamingo uses industry sector classification as an intermediate layer to identify when there are no good matches in the database. On a dataset of 664 products, Flamingo achieves an EIF matching precision of 75%.

AtmoRep is a novel, task-independent stochastic computer model of atmospheric dynamics inspired by the concept of foundation models in natural language processing, like the GPT line or PalmX, applied in the context of Earth system science. The main innovative aspect consists in the fact that the model can skillfully solve scientific tasks it was not specifically trained on, clearly exhibiting in-context learning capabilities. AtmoRep's skill has been tested on nowcasting, temporal interpolation, model correction, and counterfactuals, demonstrating that large-scale neural networks can provide skillful, task-independent models able to complement the existing numerical approaches in multiple applications. In addition, the authors also demonstrated the possibility to further increase the model accuracy by fine tuning it directly on observational data for tasks such as precipitation corrections or downscaling.

The oil and gas sector is the second largest anthropogenic emitter of methane, which is responsible for approximately 25% of global warming since pre-industrial times. In order to mitigate methane atmospheric emissions from oil and gas industry, the potential emitting infrastructure must be monitored. Initiatives such as the Methane Alert and Response System (MARS), launched by the United Nations Environment Program, aim to locate significant emissions events, alert relevant stakeholders, as well as monitor and track progress in mitigation efforts. To achieve this goal, an automated solution is needed for consistent monitoring across multiple oil and gas basins around the world. Most methane emissions analysis studies propose post-emission analysis. The works and future guidelines presented in this paper aim to provide an automated collection of informed methane emissions by oil and gas site and infrastructure which are necessary to dress emission profile in near real time. This proposed framework also permits to create action margins to reduce methane emissions by passing from post methane emissions analysis to forecasting methods.

Fusing abundant satellite data with sparse ground measurements constitutes a major challenge in climate modeling. To address this, we propose a strategy to augment the training dataset by introducing unlabeled satellite images paired with pseudo-labels generated through a spatial interpolation technique known as ordinary kriging, thereby making full use of the available satellite data resources. We show that the proposed data augmentation strategy helps enhance the performance of the state-of-the-art convolutional neural network-random forest (CNN-RF) model by a reasonable amount, resulting in a noteworthy improvement in spatial correlation and a reduction in prediction error.

Photovoltaic (PV) energy is crucial for the decarbonization of energy systems. Due to the lack of centralized data, remote sensing of rooftop PV installations is the best option to monitor the evolution of the rooftop PV installed fleet at a regional scale. However, current techniques lack reliability and are notably sensitive to shifts in the acquisition conditions. To overcome this, we leverage the wavelet scale attribution method (WCAM), which decomposes a model's prediction in the space-scale domain. The WCAM enables us to assess on which scales the representation of a PV model rests and provides insights to derive methods that improve the robustness to acquisition conditions, thus increasing trust in deep learning systems to encourage their use for the safe integration of clean energy in electric systems.

While deep learning gradually penetrates operational planning of power systems, its inherent prediction errors may significantly affect electricity prices. This paper examines how prediction errors propagate into electricity prices, revealing notable pricing errors and their spatial disparity in congested power systems. To improve fairness, we propose to embed electricity market-clearing optimization as a deep learning layer. Differentiating through this layer allows for balancing between prediction and pricing errors, as oppose to minimizing prediction errors alone. This layer implicitly optimizes fairness and controls the spatial distribution of price errors across the system. We showcase the price-aware deep learning in the nexus of wind power forecasting and short-term electricity market clearing.

Deforestation is a major driver of climate change. To mitigate deforestation, carbon offset projects aim to protect forest areas at risk. However, existing literature shows that most projects have substantially overestimated the risk of deforestation, thereby issuing carbon credits without equivalent emissions reductions. In this study, we examine if the spread of deforestation can be predicted ex-ante using Deep Learning (DL) models. Our input data includes past deforestation development, slope information, land use, and other terrain- and soil-specific covariates. Testing predictions 1-year ahead, we find that our models only achieve low levels of predictability. For pixel-wise classification at a 30 m resolution, our models achieve an F1 score of 0.263. Only when substantially simplifying the task to predicting if any level of deforestation occurs within a 1.5 km squared tile, the model results improve to a moderate performance (F1: 0.608). We conclude that, based on our input data, deforestation cannot be predicted accurately enough to justify the ex-ante issuance of carbon credits for forest conservation projects. As main challenges, there is the extreme class imbalance between pixels that are deforested (minority) and not deforested (majority) as well as the omittance of social, political, and economic drivers of deforestation.

Emulators, or reduced-complexity models, serve as an ideal complement to earth system models (ESM) by providing the climate information under various scenarios at much lower computational costs. We develop an emulator of climate extremes that produce the temporal evolution of probability distributions of local variables on a spatially resolved grid. The representative modes of climate change are identified using principal component analysis (PCA), and the PCA time series are approximated using stochastic models. When applied to ERA5 data, the model accurately reproduces the quantiles of local daily maximum temperature and effectively captures the non-Gaussian statistics. We also discuss potential generalization of our emulator to different climate change scenarios.

Airborne Wind Energy (AWE) is an emerging technology that promises to be able to harvest energy from strong high-altitude winds, while addressing some of the key critical issues of current wind turbines. AWE is based on flying devices (usually gliders or kites) that, tethered to a ground station, fly driven by the wind and convert the mechanical energy of wind into electrical energy by means of a generator.Such systems are usually controlled by adjusting the trajectory of the kite using optimal control techniques, such as model-predictive control. These methods are based upon a mathematical model of the system to control, and they produce results that are strongly dependent on the specific model at use and difficult to generalize. Our aim is to replace these classical techniques with an approach based on Reinforcement Learning (RL), which can be used even in absence of a known model. Experimental results prove that RL is a viable method to control AWE systems in complex simulated environments, including turbulent flows.

Due to the rapidly changing climate, the frequency and severity of extreme weather, such as storms and heatwaves is expected to increase drastically over the coming decades. Accurately quantifying the risk of such events with high spatial resolution is a critical step in the implementation of strategies to prepare for and mitigate the damages. As fully resolved simulations remain computationally out of reach, policy makers must rely on coarse resolution climate models which either parameterize or completely ignore sub-grid scale dynamics. In this work we propose a machine learning framework to debias under-resolved simulations of complex and chaotic dynamical systems such as atmospheric dynamics. The proposed strategy uses ``nudged'' simulations of the coarse model to generate training data designed to minimize the effects of chaotic divergence. We illustrate through a prototype QG model that the proposed approach allows us to machine learn a map from the chaotic attractor of under-resolved dynamics to that of the fully resolved system. In this way we are able to recover extreme event statistics using a very small training dataset.

Improving the accuracy of sub-seasonal to seasonal (S2S) extremes can significantly impact society. Providing S2S forecasts in risk or extreme indices can aid disaster response, especially for drought and flood events. Additionally, it can provide updates on disease outbreaks and aid in predicting the occurrence, duration, and decline of heat waves. This work uses a transformer model to predict the daily temperature distributions in the S2S scale. We analyze how the model performs in extreme temperatures by comparing its output distributions with those obtained from ECMWF forecasts across different metrics. Our model produces better responses for temperatures in average and extreme regions. Also, we show how our model better captures the heatwave that hit Europe in the summer of 2019.

Thermal Computational Fluid Dynamics (CFD) models analyze airflow and heat distribution in data centers, but their complex computations hinder efficient energy-saving optimizations for sustainability. We introduce a new method to acquire data and model 3D Convolutional Neural Network (CNN) based surrogates for CFDs, which predict a data center's temperature distribution based on server workload, HVAC airflow rate, and temperature set points. The surrogate model's predictions are highly accurate, with a mean absolute error of 0.31°C compared to CFD-based ground truth temperatures. The surrogate model is three orders of magnitude faster than CFDs in generating the temperature maps for similar-sized data centers, enabling real-time applications. It helps to quickly identify and reduce temperature hot spots($7.7%) by redistributing workloads and saving cooling energy($2.5%). It also aids in optimizing server placement during installation, preventing issues, and increasing equipment lifespan. These optimizations boost sustainability by reducing energy use, improving server performance, and lowering environmental impact.

There have been growing discussions on estimating and subsequently reducing the operational carbon footprint of enterprise data centers. The design and intelligent control for data centers have an important impact on data center carbon footprint. In this paper, we showcase PyDCM, a Python library that enables extremely fast prototyping of data center design and applies reinforcement learning-enabled control with the purpose of evaluating key sustainability metrics, including carbon footprint, energy consumption, and observing temperature hotspots. We demonstrate these capabilities of PyDCM and compare them to existing works in EnergyPlus for modeling data centers. PyDCM can also be used as a standalone Gymnasium environment for demonstrating sustainability-focused data center control.

Multiple studies have now demonstrated that machine learning (ML) can give improved skill for predicting or simulating fairly typical weather events, for tasks such as short-term and seasonal weather forecasting, downscaling simulations to higher resolution and emulating and speeding up expensive model parameterisations. Many of these used ML methods with very high numbers of parameters, such as neural networks, which are the focus of the discussion here. Not much attention has been given to the performance of these methods for extreme event severities of relevance for many critical weather and climate prediction applications, with return periods of more than a few years. This leaves a lot of uncertainty about the usefulness of these methods, particularly for general purpose prediction systems that must perform reliably in extreme situations. ML models may be expected to struggle to predict extremes due to there usually being few samples of such events. However, there are some studies that do indicate that ML models can have reasonable skill for extreme weather, and that it is not hopeless to use them in situations requiring extrapolation. This article reviews these studies and argues that this is an area that needs researching more. Ways to get a better understanding of how well ML models perform at predicting extreme weather events are discussed.

Climate change has resulted in a year over year increase in adverse weather and weather conditions which contribute to increasingly severe fire seasons. These fires threaten life, property, ecology, cultural heritage, and critical infrastructure. In this paper, we introduce SimFire, a versatile wildland fire projection simulator designed to generate realistic wildfire scenarios, and SimHarness, a modular agent-based machine learning wrapper capable of automatically generating land management strategies within SimFire to reduce the overall damage to the area. Together, this publicly available system allows researchers and practitioners the ability to emulate and assess the effectiveness of firefighter interventions and formulate strategic plans that prioritize value preservation and resource allocation optimization.

Everyday citizens, emergency responders, and critical infrastructure can be dramatically affected by the flooding of streets and roads. Climate change exacerbates these floods through sea level rise and more frequent major storm events. Low-level flooding, such as nuisance flooding, continues to increase in frequency, especially in cities like Norfolk, Virginia, which can expect nearly 200 flooding events by 2050 [1]. Recently, machine learning (ML) models have been leveraged to produce real-time predictions based on local weather and geographic conditions. However, ML models are known to produce unusual results when presented with data that varies from their training set. For decision-makers to determine the trustworthiness of the model's predictions, ML models need to quantify their prediction uncertainty. This study applies Deep Quantile Regression to a previously published, Long Short-Term Memory-based model for hourly water depth predictions [2], and analyzes its out-of-distribution performance.

Deep learning for detecting objects in remotely sensed imagery can enable new technologies for important applications including mitigating climate change. However, these models often require large datasets labeled with bounding box annotations which are expensive to curate, prohibiting the development of models for new tasks and geographies. To address this challenge, we develop weakly-semi-supervised object detection (WSSOD) models on remotely sensed imagery which can leverage a small amount of bounding boxes together with a large amount of point labels that are easy to acquire at scale in geospatial data. We train WSSOD models which use large amounts of point-labeled images with varying fractions of bounding box labeled images in FAIR1M and a wind turbine detection dataset, and demonstrate that they substantially outperform fully supervised models trained with the same amount of bounding box labeled images on both datasets. Furthermore, we find that the WSSOD models trained with 2-10x fewer bounding box labeled images can perform similarly to or outperform fully supervised models trained on the full set of bounding-box labeled images.We believe that the approach can be extended to other remote sensing tasks to reduce reliance on bounding box labels and increase development of models for impactful applications.

While fusion reactors known as tokamaks hold promise as a firm energy source,advances in plasma control, and handling of events where control of plasmas is lost, are needed for them to be economical. A significant bottleneck towards applying more advanced control algorithms is the need for better plasma simulation, where both physics-based and data-driven approaches currently fall short. The former is bottle-necked by both computational cost and the difficulty of modelling plasmas, and the latter is bottle-necked by the relative paucity of data. To address this issue, this work applies the neural ordinary differential equations (ODE) framework to the problem of predicting a subset of plasma dynamics, namely the coupled plasma current and internal inductance dynamics. As the neural ODE framework allows for the natural inclusion of physics-based inductive biases, we train both physics-based and neural network models on data from the Alcator C-Mod fusion reactor and find that a model that combines physics-based equations with a neural ODE performs better than both existing physics-motivated ODEs and a pure neural ODE model.

In global climate models, small-scale physical processes are represented using subgrid-scale (SGS) models known as parameterizations, and these parameterizations contribute substantially to uncertainties in climate projections. Recently, machine learning techniques, particularly deep neural networks (NNs), have emerged as novel tools for developing SGS parameterizations. Different strategies exist for training these NN-based SGS models. Here, we use a 1D model of the quasi-biennial oscillation (QBO) and atmospheric gravity wave (GW) parameterizations as testbeds to explore various learning strategies and challenges due to scarcity of high-fidelity training data. We show that a 12-layer convolutional NN that predicts GW forcings for given wind profiles, when trained offline in a big-data regime (100-years), produces realistic QBOs once coupled to the 1D model. In contrast, offline training of this NN in a small-data regime (18-months) yields unrealistic QBOs. However, online re-training of just two layers of this NN using ensemble Kalman inversion and only time-averaged QBO statistics leads to parameterizations that yield realistic QBOs. Fourier analysis of these three NNs’ kernels suggests how/why re-training works and reveals that these NNs primarily learn low-pass, high-pass, and a combination of band-pass Gabor filters, consistent with the importance of both local and non-local dynamics in GW propagation/dissipation. These strategies/findings apply to data-driven parameterizations of other climate processes generally.

The Madden–Julian Oscillation (MJO) is an influential climate phenomenon that plays a vital role in modulating global weather patterns. In spite of the improvement in MJO predictions made by machine learning algorithms, such as neural networks, most of them cannot provide the uncertainty levels in the MJO forecasts directly. To address this problem, we develop a nonparametric strategy based on Gaussian process (GP) models. We calibrate GPs using empirical correlations. Furthermore, we propose a posteriori covariance correction that extends the probabilistic coverage by more than three weeks.

Comprehensive information on housing stock is crucial for climate adaptation initiatives aiming to reduce the adverse impacts of climate-extreme hazards in high-risk regions like the Caribbean. In this study, we propose a workflow for rapidly generating critical baseline housing stock data using very high-resolution drone images and deep learning techniques. Specifically, our work leverages the Segment Anything Model and convolutional neural networks for the automated generation of building footprint and roof classification maps. By enhancing local capacity in government agencies, this work seeks to improve the climate resilience of the housing sector in small island developing states in the Caribbean.

Detecting temporal shifts in plant flowering times is of increasing importance in a context of climate change, with applications in plant ecology, but also health, agriculture, and ecosystem management. However, scaling up plant-level monitoring is cost prohibitive, and flowering transitions are complex and difficult to model. We develop two sets of approaches to detect the onset of flowering at large-scale and high-resolution. Using fine grain temperature data with domain knowledge based features, and traditional machine learning models provides the best performance. Using satellite data, with deep learning to deal with high dimensionality and transfer learning to overcome ground truth label sparsity, is also a promising approach, as it reaches good performance with more systematically available data.

Studying glacier movements is crucial because of their indications for global climate change and its effects on local land masses. Building on established methods for glacier movement prediction from Landsat-8 satellite imaging data, we develop an attention-based deep learning model for time series data prediction of glacier movements. In our approach, the Normalized Difference Snow Index is calculated from the Landsat-8 spectral reflectance bands for data of the Parvati Glacier (India) to quantify snow and ice in the scene images, which is then used for time series prediction. Based on this data, a newly developed Long-Short Term Memory Encoder-decoder neural network model is trained, incorporating a Multi-head Self Attention mechanism in the decoder. The model shows promising results, making the prediction of optical flow vectors from pure model predictions possible.

Complex ocean systems such as the Antarctic Circumpolar Current play key roles in the climate, and current models predict shifts in their strength and area under climate change. However, the physical processes underlying these changes are not well understood, in part due to the difficulty of characterizing and tracking changes in ocean physics in complex models. To understand changes in the Antarctic Circumpolar Current, we extend the method Tracking global Heating with Ocean Regimes (THOR) to a mesoscale eddy permitting climate model and identify regions of the ocean characterized by similar physics, called dynamical regimes, using readily accessible fields from climate models. To this end, we cluster grid cells into dynamical regimes and train an ensemble of neural networks to predict these regimes and track them under climate change. Finally, we leverage this new knowledge to elucidate the dynamics of regime shifts. Here we illustrate the value of this high-resolution version of THOR, which allows for mesoscale turbulence, with a case study of the Antarctic Circumpolar Current and its interactions with the Pacific-Antarctic Ridge. In this region, THOR specifically reveals a shift in dynamical regime under climate change driven by changes in wind stress and interactions with bathymetry. Using this knowledge to guide further exploration, we find that as the Antarctic Circumpolar Current shifts north under intensifying wind stress, the dominant dynamical role of bathymetry weakens and the flow strengthens.

Climate change is increasing the intensity and frequency of many extreme weather events, including heatwaves, which results in increased thermal discomfort and mortality rates. While global mitigation action is undoubtedly necessary, so is climate adaptation, e.g., through climate-sensitive urban planning. Among the most promising strategies is harnessing the benefits of urban trees in shading and cooling pedestrian-level environments. Our work investigates the challenge of optimal placement of such trees. Physical simulations can estimate the radiative and thermal impact of trees on human thermal comfort but induce high computational costs. This rules out optimization of tree placements over large areas and considering effects over longer time scales. Hence, we employ neural networks to simulate the point-wise mean radiant temperatures--a driving factor of outdoor human thermal comfort--across various time scales, spanning from daily variations to extended time scales of heatwave events and even decades. To optimize tree placements, we harness the innate local effect of trees within the iterated local search framework with tailored adaptations. We show the efficacy of our approach across a wide spectrum of study areas and time scales. We believe that our approach is a step towards empowering decision-makers, urban designers and planners to proactively and effectively assess the potential of urban trees to mitigate heat stress.

For low carbon transport planning it's essential to grasp the location-specific cause-and-effect mechanisms that the built environment has on travel. Yet, current research falls short in representing causal relationships between the "6D" urban form variables and travel, generalizing across different regions, and modelling urban form effects at high spatial resolution. Here, we address these gaps by utilizing a causal discovery and an explainable machine learning framework to detect urban form effects on intra-city travel emissions based on high-resolution mobility data of six cities across three continents. We show that distance to center, demographics and density indirectly affect other urban form features and that location-specific influences align across cities, yet vary in magnitude. In addition, the spread of the city and the coverage of jobs across the city are the strongest determinants of travel-related emissions, highlighting the benefits of compact development and associated benefits. Our work is a starting point for location-specific analysis of urban form effects on mobility using causal discovery approaches, which is highly relevant municipalities across continents.

Glacier retreat is a key indicator of climate change and requires regular updates of the glacier area. Recently, the release of a new inventory for the European Alps showed that glaciers continued to retreat at about 1.3% per year from 2003 to 2015. The outlines were produced by manually correcting the results of a semi-automatic method applied to Sentinel-2 imagery. In this work we develop a fully-automatic pipeline based on Deep Learning to investigate the evolution of the glaciers in the Alps from 2015 to present (2023). After outlier filtering, we provide individual estimates for around 1300 glaciers, representing 87% of the glacierized area. Regionally we estimate an area loss of -1.8% per year, with large variations between glaciers. Code and data are available at https://anonymous.4open.science/r/glaciermappingalps.

Understanding potential changes in future rainfall and their local impacts on Australian communities can inform adaptation decisions worth billions of dollars in insurance, agriculture, and other sectors. This understanding relies on downscaling a large ensemble of coarse Global Climate Models (GCMs), our primary tool for simulating future climate. However, the prohibitively high computational cost of downscaling has been a significant barrier. In response, this study develops a cost-efficient downscaling framework for daily precipitation using Long Short-Term Memory (LSTM) models. The models are trained with ERA5 reanalysis data and a customized quantile loss function to better capture precipitation extremes. The framework is employed to downscale precipitation from a GCM member of the CMIP6 ensemble. We demonstrate the skills of the downscaling models to capture spatial and temporal characteristics of precipitation. We also explore regional future changes in precipitation extremes projected by the downscaled GCM. In general, this framework will enable the generation of a large ensemble of regional future projections for Australian rainfall. This will further enhance the assessment of likely climate risks and the quantification of their uncertainties.

Climate hazards are spatial phenomena that 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 to extreme values. Once trained, the model can be used to efficiently generate spatial compound hazard events, which are urgently needed for climate disaster preparedness. The method developed is flexible and transferable to other multivariate and spatial climate datasets.

The rise of accurate machine learning methods for weather forecasting is creating radical new possibilities for modeling the atmosphere. In the time of climate change, having access to high-resolution forecasts from models like these is also becoming increasingly vital. While most existing Neural Weather Prediction (NeurWP) methods focus on global forecasting, an important question is how these techniques can be applied to limited area modeling. In this work we adapt the graph-based NeurWP approach to the limited area setting and propose a multi-scale hierarchical model extension. Our approach is validated by experiments with a local model for the Nordic region.

Coastal zones are under increasing pressures due to climate change and the increasing population densities in coastal areas around the globe. Our ability to accurately forecast the evolution of the coastal zone is of critical importance to coastal managers in the context of risk assessment and mitigation. Recent advances in artificial intelligence and remote sensing enable the development of automatic large-scale analysis methodologies based on observation data. In this work, we make use of a novel satellite-derived shoreline forecasting dataset and a variant of the common Encoder-Decoder neural network, UNet, in order to predict shoreline change based on spatio-temporal data. We analyze the importance of including the spatial context at the prediction step and we find that it greatly enhances model performance. Overall, the model presented here demonstrates significant shoreline forecasting skill around the globe, achieving a global correlation of 0.77. The code, data, and trained models are available online at https://anonymous.4open.science/r/DL-spatiotemporal-shorelines.

Governments around the world aspire to ground decision-making on evidence. Many of the foundations of policy making — e.g. sensing patterns that relate to societal needs, developing evidence-based programs, forecasting potential outcomes of policy changes, and monitoring effectiveness of policy programs — have the potential to benefit from the use of large-scale datasets or simulations together with intelligent algorithms. These could, if designed and deployed in a way that is well grounded on scientific evidence, enable a more comprehensive, faster, and rigorous approach to policy making. Integrated Assessment Models (IAM) is a broad umbrella covering scientific models that attempt to link main features of society and economy with the biosphere into one modelling framework. At present, these systems are probed by by policy makers and advisory groups in a hypothesis-driven manner. In this paper, we empirically demonstrate that modern Reinforcement Learning can be used to probe IAMs and explore the space of solutions in a more principled manner. While the implication of our results are modest since the environment is simplistic, we believe that this is a stepping stone towards more ambitious use cases, which could allow for effective exploration of policies and understanding of their consequences and limitations.

Social media platforms such as Twitter (now known as X) have revolutionized how the public engage with important societal and political topics. Recently, climate change discussions on social media became a catalyst for political polarization and the spreading of misinformation. In this work, we aim to understand how real world events influence the opinions of individuals towards climate change related topics on social media. To this end, we extracted and analyzed a dataset of 13.6 millions tweets sent by 3.6 million users from 2006 to 2019. Then, we construct a temporal graph from the user-user mentions network and utilize the Louvain community detection algorithm to analyze the changes in community structure around Conference of the Parties on Climate Change (COP) events. Next, we also apply tools from the Natural Language Processing literature to perform sentiment analysis and topic modeling on the tweets. Our work acts as a first step towards understanding the evolution of pro-climate change communities around COP events. Answering these questions helps us understand how to raise people's awareness towards climate change thus hopefully calling on more individuals to join the collaborative effort in slowing down climate change.

Instance segmentation is pivotal for environmental sciences and climate change research, facilitating important tasks from land cover classification to glacier monitoring. This paper addresses the prevailing challenges associated with data scarcity when using traditional models like YOLOv8 by introducing a novel, data-efficient workflow for instance segmentation. The proposed Segment-then-Classify (STC) strategy leverages the zero-shot capabilities of the novel Segment Anything Model (SAM) to segment all objects in an image and then uses a simple classifier such as the Vision Transformer (ViT) to identify objects of interest thereafter. Evaluated on the VHR-10 dataset, our approach demonstrated convergence with merely 40 examples per class. YOLOv8 requires 3 times as much data to achieve the STC's peak performance. The highest performing class in the VHR-10 dataset achieved a near-perfect mAP@0.5 of 0.99 using the STC strategy. However, performance varied greatly across other classes due to the SAM model’s occasional inability to recognize all relevant objects, indicating a need for refining the zero-shot segmentation step. The STC workflow therefore holds promise for advancing few-shot learning for instance segmentation in environmental science.

The famous IPCC general circulation models (GCMs) constitute the primary tools for climate projections. Calibrating, or tuning the parameters of the models can significantly improve their predictions, thus their scientific and societal impacts. Unfortunately, traditional tuning techniques remain time-consuming and computationally costly, even at coarse resolution. A specific challenge for the tuning of climate models lies in the tuning of both fast and slow climatic features: while atmospheric processes adjust on hourly to weekly timescales, vegetation or ocean dynamics drive mechanisms of variability at decadal to millennial timescales. In this work, we explore whether and how History Matching, which uses machine learning based emulators to accelerate and automate the tuning process, is relevant for tuning climate models with multiple timescales. To facilitate this exploration, we work with a climate model of intermediate complexity, yet test experimental tuning protocols that can be directly applied to more complex GCMs to reduce uncertainty in climate projections.

An invasive species of grass known as "buffelgrass" contributes to severe wildfires and biodiversity loss in the Southwest United States. We tackle the problem of predicting buffelgrass "green-ups" (i.e. readiness for herbicidal treatment). To make our predictions, we explore temporal, visual and multi-modal models that combine satellite sensing and deep learning. We find that all of our neural-based approaches improve over conventional buffelgrass green-up models, and discuss how neural model deployment promises significant resource savings.

Global climate models (GCMs) are the main tools for understanding and predicting climate change. However, due to limited numerical resolutions, these models suffer from major structural uncertainties; e.g., they cannot resolve critical processes such as small-scale eddies in atmospheric and oceanic turbulence. Thus, such small-scale processes have to be represented as a function of the resolved scales via closures (parametrization). The accuracy of these closures is particularly important for capturing climate extremes. Traditionally, such closures are based on heuristics and simplifying assumptions about the unresolved physics. Recently, supervised-learned closures, trained offline on high-fidelity data, have been shown to outperform the classical physics-based closures. However, this approach requires a significant amount of high-fidelity training data and can also lead to instabilities. Reinforcement learning is emerging as a potent alternative for developing such closures as it requires only low-order statistics and leads to stable closures. In Scientific Multi-Agent Reinforcement Learning (SMARL) computational elements serve a dual role of discretization points and learning agents. Here, we leverage SMARL and fundamentals of turbulence physics to learn closures for canonical prototypes of atmospheric and oceanic turbulence. The policy is trained using only the enstrophy spectrum, which is nearly invariant and can be estimated from a few high-fidelity samples. We show that these closures lead to stable low-resolution simulations that, at a fraction of the cost, can reproduce the high-fidelity simulations' statistics, including the tails of the probability density functions (PDFs). These results demonstrate the high potential of SMARL for closure modeling for GCMs, especially in the regime of scarce data and indirect observations.

In recent years, the convergence of data-driven machine learning models with Data Assimilation (DA) offers a promising avenue for enhancing weather forecasting. This study delves into this emerging trend, presenting our methodologies and outcomes. We harnessed the UK's local \verb+ERA5+ 850 hPa temperature data and refined the U-STN12 global weather forecasting model, tailoring its predictions to the UK's climate nuances. From the \verb+ASOS+ network, we sourced \verb+t2m+ data, representing ground observations across the UK. We employed the advanced kriging method with a polynomial drift term for consistent spatial resolution. Furthermore, Gaussian noise was superimposed on the \verb+ERA5+ \verb+T850+ data, setting the stage for ensuing multi-time step virtual observations. Probing into the assimilation impacts, the \verb+ASOS+ \verb+t2m+ data was integrated with the \verb+ERA5+ \verb+T850+ dataset. Our insights reveal that while global forecast models can adapt to specific regions, incorporating atmospheric data in DA significantly bolsters model accuracy. Conversely, the direct assimilation of surface temperature data tends to mitigate this enhancement, tempering the model's predictive prowess.

Accurate and computationally-viable representations of clouds and turbulence are a long-standing challenge for climate model development. Traditional parameterizations that crudely but efficiently approximate these processes are a leading source of uncertainty in long-term projected warming and precipitation patterns. Machine Learning (ML)-based parameterizations have long been hailed as a promising alternative with the potential to yield higher accuracy at a fraction of the cost of more explicit simulations. However, these ML variants are often unpredictably unstable and inaccurate in online testing (i.e. in a downstream hybrid simulation task where they are dynamically coupled to the large-scale climate model). These issues are exacerbated in out-of-distribution climates. Certain design decisions such as ``climate-invariant" feature transformation, input vector expansion, and temporal history incorporation have been shown to improve online performance, but they may be insufficient for the mission-critical task of online out-of-distribution generalization. If feature selection and transformations can inoculate hybrid physics-ML climate models from non-physical out-of-distribution extrapolation in a changing climate, there is far greater potential in extrapolating from observational data. Otherwise, training on multiple simulated climates becomes an inevitable necessity. While our results show generalization benefits from these design decisions, such benefits do not sufficiently preclude the necessity of using multi-climate simulated training data.

Building operations account for a significant portion of global emissions, contributing approximately 28\% of global greenhouse gas emissions. With anticipated increase in cooling demand due to rising global temperatures, the optimization of rooftop units (RTUs) in buildings becomes crucial for reducing emissions. We focus on the optimization of the economizer logic within RTUs, which balances the mix of indoor and outdoor air. By effectively utilizing free outside air, economizers can significantly decrease mechanical energy usage, leading to reduced energy costs and emissions. We introduce a reinforcement learning (RL) approach that adaptively controls the economizer based on the unique characteristics of individual facilities. We have trained and deployed our solution in the real-world across a distributed building stock. We address the scaling challenges with our cloud-based RL deployment on 10K+ RTUs across 200+ sites.

Floods can cause horrific harm to life and property. However, they can be mitigated or even avoided by the effective use of hydraulic structures such as dams, gates, and pumps. By pre-releasing water via these structures in advance of extreme weather events, water levels are sufficiently lowered to prevent floods. In this work, we propose FIDLAR, a Forecast Informed Deep Learning Architecture, achieving flood management in watersheds with hydraulic structures in an optimal manner by balancing out flood mitigation and unnecessary wastage of water via pre-releases. We perform experiments with FIDLAR using data from the South Florida Water Management District, which manages a coastal area that is highly prone to frequent storms and floods. Results show that FIDLAR performs better than the current state-of-the-art with several orders of magnitude speedup and with provably better pre-release schedules. The dramatic speedups make it possible for FIDLAR to be used for real-time flood management. The main contribution of this paper is the effective use of tools for model explainability, allowing us to understand the contribution of the various environmental factors towards its decisions.

The rapid growth of machine learning (ML) has led to an increased demand for computational power, resulting in larger data centers (DCs) and higher energy consumption. To address this issue and reduce carbon emissions, intelligent control of DC components such as cooling, load shifting, and energy storage is essential. However, the complexity of managing these controls in tandem with external factors like weather and green energy availability presents a significant challenge. While some individual components like HVAC control have seen research in Reinforcement Learning (RL), there's a gap in holistic optimization covering all elements simultaneously. To tackle this, we've developed DCRL, a multi-agent RL environment that empowers the ML community to research, develop, and refine RL controllers for carbon footprint reduction in DCs. DCRL is a flexible, modular, scalable, and configurable platform that can handle large High Performance Computing (HPC) clusters. In its default setup, DCRL also provides a benchmark for evaluating multi-agent RL algorithms, facilitating collaboration and progress in green computing research.

Bird, insect, and other wild animal populations are rapidly declining, highlighting the need for better monitoring, understanding, and protection of Earth’s remaining wild places. However, direct monitoring of biodiversity is difficult. Passive Acoustic Monitoring (PAM) enables detection of the vocalizing species in an ecosystem, many of which can be difficult or impossible to detect by satellite or camera trap. Large-scale PAM deployments using low-cost devices allow measuring changes over time and responses to environmental changes, and targeted deployments can discover and monitor endangered or invasive species. Machine learning methods are needed to analyze the thousands or even millions of hours of audio produced by large-scale deployments. But there are a massive number of potential signals to target for bioacoustic measurement, and many of the most interesting lack training data. Many rare species are difficult to observe. Detecting specific call-types and juvenile calls can give further insight into behavior and population health, but almost no structured datasets exist for these use-cases. No single classifier can address all of these needs, so practitioners regularly need to create new classifiers to address novel problems. Soundscape annotation efforts are very expensive, and machine learning experts are scarce, creating a bottleneck on analysis. We aim to eliminate the bottleneck by providing an efficient, self-contained active learning workflow for biologists.In this tutorial, we present an integrated workflow for analyzing large unlabeled bioacoustic datasets, adapting new agile modeling techniques to audio. Our goal is to allow experts to create a new high quality classifier for a novel class with under one hour of effort. We achieve this by leveraging transfer learning from high-quality bioacoustic models, vector search over audio databases, and lightweight Python notebook UX. The workflow can begin from a single example, proceeds through an efficient active learning loop, and finally applies the produced classifier to a large mass of unlabeled data to produce insights for ecologists and land managers.

Urban Building Energy Modeling (UBEM) is an emerging method to investigate urban design and energy systems against the increasing energy demand at urban and neighborhood levels. However, current UBEM methods are mostly physic-based and time-consuming in multiple climate change scenarios. This work proposes CityTFT, a data-driven UBEM framework, to accurately model the energy demands in urban environments. With the empowerment of the underlying TFT framework and an augmented loss function, CityTFT could predict heating and cooling triggers in unseen climate dynamics with an F1 score of 99.98 \% while RMSE of loads of 13571.3750 Wh.

Renewable energy plays a crucial role in mitigating climate change. With the rising use of distributed energy resources (DERs), microgrids (MGs) have emerged as a solution to accommodate high DER penetration. However, controlling MGs' voltage during islanded operation is challenging due to system's nonlinearity and stochasticity. Although multi-agent reinforcement learning (MARL) methods have been applied to distributed MG voltage control, they suffer from bad scalability and are found difficult to control the MG with a large number of DGs due to the well-known curse of dimensionality. To address this, we propose a scalable network-aware reinforcement learning framework which exploits network structure to truncate the critic's Q-function to achieve scalability. Our experiments show effective control of a MG with up to 84 DGs, surpassing the existing maximum of 40 agents in the existing literature. We also compare our framework with state-of-the-art MARL algorithms to show the superior scalability of our framework.

This paper explores the integration of reinforcement learning (RL) into transportation simulations to explore system interventions to reduce greenhouse gas emissions. The study leverages the Behavior, Energy, Automation, and Mobility (BEAM) transportation simulation framework in conjunction with the Berkeley Integrated System for Transportation Optimization (BISTRO) for scenario development. The main objective is to determine optimal parameters for transportation simulations to increase public transport usage and reduce individual vehicle reliance. Initial experiments were conducted on a simplified transportation scenario, and results indicate that RL can effectively find system interventions that increase public transit usage and decrease transportation emissions.

Grid decarbonization for climate change requires dispatchable carbon-free energy like nuclear fusion. The tokamak concept offers a promising path for fusion, but one of the foremost challenges in implementation is the occurrence of energetic plasma disruptions. In this study, we delve into Machine Learning approaches to predict plasma state outcomes. Our contributions are twofold: (1) We present a novel application of Continuous Convolutional Neural Networks for disruption prediction and (2) We examine the advantages and disadvantages of continuous models over discrete models for disruption prediction by comparing our model with the previous, discrete state of the art, and show that continuous models offer significantly better performance (Area Under the Receiver Operating Characteristic Curve = 0.974 v.s. 0.799) with fewer parameters.

Understanding the future climate is crucial for informed policy decisions on climatechange prevention and mitigation. Earth system models play an important rolein predicting future climate, requiring accurate representation of complex sub-processes that span multiple time scales and spatial scales. One such process thatlinks seasonal and interannual climate variability to cyclical biological events istree phenology in deciduous forests. Phenological dates, such as the start andend of the growing season, are critical for understanding the exchange of carbonand water between the biosphere and the atmosphere. Mechanistic predictionof these dates is challenging. Hybrid modelling, which integrates data-drivenapproaches into complex models, offers a solution. In this work, as a first steptowards this goal, train a deep neural network to predict a phenological index frommeteorological time series. We find that this approach outperforms traditionalprocess-based models. This highlights the potential of data-driven methods toimprove climate predictions. We also analyze which variables and aspects of thetime series influence the predicted onset of the season, in order to gain a betterunderstanding of the advantages and limitations of our model.

Predicting typhoon intensity accurately across space and time is crucial for issuing timely disaster warnings and facilitating emergency response. This has vast potential for minimizing the loss of life, property damage, and reducing economic and environmental impacts. Leveraging satellite imagery for situational analysis is effective but presents challenges due to the complex relationships among clouds and the highly dynamic context. Existing deep learning methods in this domain rely on convolutional neural networks (CNNs), which suffer from limited per-layer receptive fields. This limitation hinders their ability to capture long-range dependencies and global contextual knowledge during inference. In response, we introduce a novel approach, the "Typhoon Intensity Transformer" (TiT), which leverages self-attention mechanisms with global receptive fields per layer. TiT adopts a sequence-to-sequence feature representation learning perspective. It begins by dividing a given satellite image into a sequence of patches and recursively employs self-attention operations to extract both local and global contextual relationships between all patch pairs simultaneously, enhancing per-patch feature representation learning. Extensive experiments on a publicly available typhoon benchmark validate the efficacy of TiT when compared to both state-of-the-art deep learning methods and conventional meteorological modeling approaches.

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.

Inverse modelling systems relying on Lagrangian Particle Dispersion Models (LPDMs) are a popular way to quantify greenhouse gas (GHG) emissions using atmospheric observations, providing independent validation to countries' self-reported emissions. However, the increased volume of satellite measurements cannot be fully leveraged due to computational bottlenecks. Here, we propose a data-driven architecture with Graph Neural Networks that emulates the outputs of LPDMs using only meteorological inputs, and demonstrate it in application with preliminary results for satellite measurements over Brazil.

Monitoring air pollution is of vital importance to the overall health of the population. Unfortunately, devices that can measure air quality can be expensive, and many cities in low and middle-income countries have to rely on a sparse allocation of them. In this paper, we investigate the use of Gaussian Processes for both nowcasting the current air-pollution in places where there are no sensors and forecasting the air-pollution in the future at the sensor locations. In particular, we focus on the city of Kampala in Uganda, using data from AirQo's network of sensors. We demonstrate the advantage of removing outliers, compare different kernel functions and additional inputs. We also compare two sparse approximations to allow for the large amounts of temporal data in the dataset.

Machine learning (ML)-based weather models have recently undergone rapid improvements.These models are typically trained on gridded reanalysis data from numerical data assimilation systems. However, reanalysis data comes with limitations, such as assumptions about physical laws and low spatiotemporal resolution. The gap between reanalysis and reality has sparked growing interest in training ML models directly on observations such as weather stations. Modelling scattered and sparse environmental observations requires scalable and flexible ML architectures, one of which is the convolutional conditional neural process (ConvCNP). ConvCNPs can learn to condition on both gridded and off-the-grid context data to make uncertainty-aware predictions at target locations. However, the sparsity of real observations presents a challenge for data-hungry deep learning models like the ConvCNP. One potential solution is `Sim2Real': pre-training on reanalysis and fine-tuning on observational data. We analyse Sim2Real with a ConvCNP trained to interpolate surface air temperature over Germany, using varying numbers of weather stations for fine-tuning. On held-out weather stations, Sim2Real training substantially outperforms the same model trained only with reanalysis data or only with station data, showing that reanalysis data can serve as a stepping stone for learning from real observations. Sim2Real could enable more accurate models for climate change monitoring and adaptation.

Methane emissions have a significant impact on increasing global warming. Satellite-based methane detection methods can help mitigate methane emissions, as they provide a constant and global detection. The Sentinel-2 constellation, in particular, offers frequent and publicly accessible images on a global scale. We propose a deep learning approach to detect methane plumes from Sentinel-2 images. We construct a dataset of 5200 satellite images with identified methane plumes, on which we train a U-Net model. Preliminary results demonstrate that the model is able to correctly identify methane plumes on training data, although generalization to new methane plumes remains challenging. All code, data, and models are made available online.

Methane (CH4) possesses a notably higher warming potential than carbon dioxide despite its lower atmospheric concentration, making it integral to global climate dynamics. Wetlands stand out as the predominant natural contributor to global methane emissions. Accurate modeling of methane emissions from wetlands is crucial for understanding and predicting climate change dynamics. However, such modeling efforts are often constrained by the inherent uncertainties in model parameters. Our work leverages machine learning (ML) to calibrate five physical parameters of the Energy Exascale Earth System Model (E3SM) land model (ELM) to improve the model’s accuracy in simulating wetland methane emissions. Unlike traditional deterministic calibration methods that target a single set of optimal values for each parameter, Bayesian calibration takes a probabilistic approach and enables capturing the inherent uncertainties in complex systems and providing robust parameter distributions for reliable predictions. However, Bayesian calibration requires numerous model runs and makes it computationally expensive. We employed an ML algorithm, Gaussian process regression (GPR), to emulate the ELM’s methane model, which dramatically reduced the computational time from 6 CPU hours to just 0.72 milliseconds per simulation. We exemplified the procedure at a representative FLUXNET-CH4 site (US-PFa) with the longest continuous methane emission data. Results showed that the default values for two of the five parameters examined were not aligned well with their respective posterior distributions, suggesting that the model’s default parameter values might not always be optimal for all sites, and that site-specific analysis is warranted. In particular, analyses at sites with different vegetation types and wetland characteristics could reveal more useful insights for understanding methane emissions modeling.

Climate models, such as Earth system models (ESMs), are crucial for simulating future climate change based on projected Shared Socioeconomic Pathways (SSP) greenhouse gas emissions scenarios. While ESMs are sophisticated and invaluable, machine learning-based emulators trained on existing simulation data can project additional climate scenarios much faster and are computationally efficient. However, they often lack generalizability and interpretability. This work delves into the potential of causal representation learning, specifically the Causal Discovery with Single-parent Decoding (CDSD) method, which could render climate model emulation efficient and interpretable. We evaluate CDSD on multiple climate datasets, focusing on emissions, temperature, and precipitation. Our findings shed light on the challenges, limitations, and promise of using CDSD as a stepping stone towards more interpretable and robust climate model emulation.

Climate change and housing growth in ecologically vulnerable areas are increasing the frequency and intensity of wildfire events. These events can have profound impact on communities and the economy, but the financial and operational impacts of wildfire on businesses have not been evaluated extensively yet. This paper presents a Wildfire Vulnerability Index (WVI) that measures the risk of a business failing in the 12 months following a wildfire event. An XGBoost algorithm champion model is compared to two challenger models: 1) a model that extends the champion model by incorporating building and property characteristics and 2) a model that utilizes a neural network approach. Results show that while all models perform well in predicting business failure risk post-disaster event, the model that uses building and property characteristics performs best across all performance metrics. As the natural environment shifts under climate change and more businesses are exposed to the consequences of wildfire, the WVI can help emergency managers allocate disaster aid to businesses at the highest risk of failing and can also provide valuable risk insights for portfolio managers and loan processors.

Wildfires are becoming unpredictable natural hazards in many regions due to climate change.However, existing state-of-the-art wildfire forecasting tools, such as the Fire Weather Index (FWI), rely solely on meteorological input parameters and have limited ability to model the increasingly dynamic nature of wildfires.In response to the escalating threat, our work addresses this shortcoming in short-term fire hazard prediction.First, we present a comprehensive and high fidelity remotely sensed active fire dataset fused from over 20 satellites.Second, we develop region-specific ML-based wildfire hazard prediction models for South America, Australia, and Southern Europe.The different models cover pixel-wise, spatial and spatio-temporal architectures, and utilize weather, fuel and location data.We evaluate the models using time-based cross-validation and can show superior performance with a PR-AUC score up to 44 times higher compared to the baseline FWI model. Using explainable AI methods, we show that these data-driven models are also capable of learning meaningful physical patterns and inferring region-specific wildfire drivers.

Cloud and precipitation properties are controlled by both local and large-scale forcings. Current weather and climate models represent clouds and precipitation through parameterizations that are based on theoretical relationships between environment, clouds, and precipitation. However, these relationships vary considerably among different weather and cloud conditions, thereby leading to inaccurate simulation of cloud and precipitation properties. In this study, we use observations from a site in the Eastern North Atlantic Ocean (28W, 39.5N) to establish a potential causal relationship between large-scale environment, cloud, and precipitation properties. We estimate the structure of a directed acyclic graph (DAG) with the NOTEARS algorithm (Non-combinatorial Optimization via Trace Exponential and Augmented lagRangian for Structure learning) (Zheng et al., 2018 \cite{Zheng2018DAGsLearning}) with a multi-layer perceptron (MLP) neural network classification architecture. We classify liquid water path (LWP), rain rate, and rain drop diameter in two classes based on lower and upper quantiles to identify the governing mechanisms responsible for the two tails of the distribution. We also invoke Random Forest classification to compare our causal model results with conventional decision tree-based approaches. We hypothesize the dominant role of cloud LWP and net radiative cooling in controlling the cloud and precipitation properties. In this way, this study demonstrates the application of a causal machine learning method to identify which environmental properties potentially control cloud and precipitation development. These results will be extremely valuable to both observational and numerical modeling communities as they could help improve the current parameterizations in the weather and climate models.

Radiative transfer modeling is a complex and computationally expensive process that is used to predict how radiation interacts with the atmosphere and Earth's surface. The Rapid Radiation Transfer Model (RRTM) is one such process model that is used in many Earth system models. In recent years, there has been a growing interest in using machine learning (ML) to speed up radiative transfer modeling. ML algorithms can be trained on large datasets of existing RRTM simulations to learn how to predict the results of new simulations without having to run the full RRTM model so one can use the algorithm for new simulations with very light computational demand. This study developed a new physics-based ML emulator for RRTM that is built on a convolutional neural network (CNN) where we trained the CNN on a dataset of 28 years of RRTM simulations. We built a custom loss function, which incorporates information on how radiation interacts with clouds at day- and night-time. The emulator was able to learn how to predict the vertical heating rates in the atmosphere with a high degree of accuracy (RMSE of less than 2% and Pearson's correlation above 0.9). The new ML emulator is over 56 times faster than the original RRTM model on traditional multi-CPU machines. This speedup could allow scientists to call the RRTM much more frequently in atmosphere models, which may improve the accuracy of climate models and reduce the uncertainty in the future climate projections.

While weather station data is a valuable resource for climate prediction, its reliability can be limited in remote locations. Furthermore, making local predictions often relies on sensor data that may not be accessible for a new, unmonitored location. In response to these challenges, we introduce a novel zero-shot learning approach designed to forecast various climate measurements at new and unmonitored locations. Our method surpasses conventional weather forecasting techniques in predicting microclimate variables by leveraging knowledge extracted from other geographic locations.

Driven by the climate crisis, the frequency and intensity of flood events are increasing. Electro-optical (EO) satellite imagery is commonly utilized for rapid disaster response. However, its utilities in flood situations are hampered by cloud cover and limited during night-time. Alternative flood detection methods utilize Synthetic Aperture Radar (SAR) data. Despite the advantages of SAR over EO in the aforementioned situations, SAR presents a distinct drawback: human analysis often struggles with data interpretation. This paper proposes a novel framework, Diffusion-based SAR-to-EO Image Translation (DSE). The DSE framework converts SAR into EO imageries, thereby enhancing the interpretability of flood insights for humans. Experimental results on the Sen1Floods11 and SEN12-FLOOD datasets confirm that the DSE framework delivers enhanced visual information and improves performance across all flood segmentation tests.

Modeling is crucial to understanding the effect of greenhouse gases, warming, and ice sheet melting on the ocean. At the same time, ocean processes affect phenomena such as hurricanes and droughts. Parameters in the models that cannot be physically measured have a significant effect on the model output. For an idealized ocean model, we generate perturbed parameter ensemble data and generate surrogate neural network models. The neural surrogates accurately predicted the one-step forward dynamics, of which we then computed the parametric sensitivity.

The potential for public misinformation fueled by “confidently wrong” Large Language Models (LLMs) is especially salient in the climate science and policy domain. We introduce the ICCS dataset, a novel, curated, expert-labeled NLP dataset consisting of 8094 climate science statements and their associated confidence levels collected from the latest IPCC AR6 reports. Using this dataset, we show that recent LLMs can classify human expert confidence in climate-related statements with reasonable—if limited—accuracy, especially in a few-shot learning setting. Overall, models exhibit consistent and significant overconfidence on low and medium confidence statements. We highlight important implications from our results for climate policy and the use of LLMs in information retrieval systems.

Glacier systems are highly sensitive to climate change and play a pivotal role in global mean sea level rise. As such, it is important to monitor how glacier velocities and ice dynamics evolve under a changing climate. The growing wealth of satellite observations has facilitated the inference of glacier velocities from remote sensing imagery through feature tracking algorithms. At present, these rely on sparse cross-correlation estimates as well as computationally expensive optical flow solutions. Here we present a novel use of deep-learning for estimating annual glacier velocities, utilizing the recurrent optical-flow based architecture, RAFT, on consecutive pairs of optical Sentinel-2 imagery. Our results highlight that deep learning can generate dense per-pixel velocity estimates within an automated framework that utilizes Sentinel-2 images over the French Alps.

Soil organic carbon (SOC) plays a pivotal role in the global carbon cycle, impacting climate dynamics and necessitating accurate estimation for sustainable land and agricultural management. While traditional methods of SOC estimation face resolution and accuracy challenges, recent advancements harness remote sensing, machine learning, and high-resolution satellite mapping. Graph Neural Networks (GNNs), especially when integrated with positional encoders, offer promise in capturing intricate relationships between soil and climate. Using the LUCAS database, this study compared four GNN operators in the positional encoder framework. Results revealed that the PESAGE and PETransformer models outperformed others in SOC estimation, indicating their potential in capturing the complex interplay between SOC and climate features. Our findings underscore the potential of GNN architectures in advancing SOC prediction, paving the way for future explorations with more advanced GNN models.

Waste segregation and recycling place a crucial role in fostering environmental sustainability. However, discerning the whether a material is recyclable or not poses a formidable challenge, primarily because of inadequate recycling guidelines to accommodate a diverse spectrum of objects and their varying conditions. We investigated the role of vision-language models in addressing this challenge. We curated a dataset consisting >1000 images across 11 disposal categories for optimal discarding and assessed the applicability of general vision-language models for recyclability classification. Our results show that Contrastive Language-Image Pre- training (CLIP) model, which is pretrained to understand the relationship between images and text, demonstrated remarkable performance in the zero-shot recycla- bility classification task, with an accuracy of 89%. Our results underscore the potential of general vision-language models in addressing real-world challenges, such as automated waste sorting, by harnessing the inherent associations between visual and textual information.

In this paper we introduce a novel framework for trajectory prediction of geospatial sequences using GraphTransformers. When viewed across several sequences, we observed that a graph structure automatically emerges between different geospatial points that is often not taken into account for such sequence modeling tasks. We show that by leveraging this graph structure explicitly, geospatial trajectory prediction can be significantly improved. Our GraphTransformer approach improves upon state-of-the-art Transformer based baseline significantly on HURDAT, a dataset where we are interested in predicting the trajectory of a hurricane on a 6 hourly basis. This helps inform evacuation efforts by narrowing down target location by 10 to 20 kilometers along both the north-south and east-west directions.

Climate change poses one of the most significant challenges to humanity. As a result of these climatic shifts,the frequency of weather, climate, and water-related disasters has multiplied fivefold over the past 50 years, resulting in over 2$ million deaths and losses exceeding U.S. $3.64 trillion.Leveraging AI-powered technologies for sustainable development and combating climate change is a promising avenue. Numerous significant publications are dedicated to using AI to improve renewable energy forecasting, enhance waste management, and monitor environmental changes in real-time. However, very few research studies focus on making AI itself environmentally sustainable. This oversight regarding the sustainability of AI within the field might be attributed to a mindset gap and the absence of comprehensive energy datasets. In addition, with the ubiquity of edge AI systems and applications, especially on-device learning, there is a pressing need to measure, analyze, and optimize their environmental sustainability, such as energy efficiency.To this end, in this paper, we propose large-scale energy datasets for edge AI, named DeepEn2023, covering a wide range of kernels, state-of-the-art deep neural network models, and popular edge AI applications.We anticipate that DeepEn2023 will enhance transparency regarding sustainability in on-device deep learning across a range of edge AI systems and applications.

This research paper delves into the unique energy challenges faced by Alaska, arising from its remote geographical location, severe climatic conditions, and heavy reliance on fossil fuels while emphasizing the shortage of comprehensive building energy data. The study introduces an ethical framework that leverages machine learning and geospatial techniques to enable the large-scale integration of data, facilitating the mapping of energy consumption data at the individual building level. Utilizing the Alaska Retrofit Information System (ARIS) and the USA Structures datasets, this framework not only identifies and acknowledges limitations inherent in existing datasets but also establishes a robust ethical foundation for data integration. This framework innovation sets a noteworthy precedent for the responsible utilization of data in the domain of climate justice research, ultimately informing the development of sustainable energy policies through an enhanced understanding of building data and advancing ongoing research agendas. Future research directions involve the incorporation of recently released datasets, which provide precise building location data, thereby further validating the proposed ethical framework and advancing efforts in addressing Alaska's intricate energy challenges.

Crop breeding is crucial in improving agricultural productivity while potentially decreasing land usage, greenhouse gas emissions, and water consumption. However, breeding programs are challenging due to long turnover times, high-dimensional decision spaces, long-term objectives, and the need to adapt to rapid climate change.This paper introduces the use of Reinforcement Learning (RL) to optimize simulated crop breeding programs. RL agents are trained to make optimal crop selection and cross-breeding decisions based on genetic information. To benchmark RL-based breeding algorithms, we introduce a suite of Gym environments.The study demonstrates the superiority of RL techniques over standard practices in terms of genetic gain when simulated in silico using real-world genomic maize data.

Land use and land cover (LULC) modeling is a challenging task due to long-range dependencies between geographic features and distinct spatial patterns related in topography, ecology, and human development. We explore the usage of a modified Pixel Constrained CNN as applied to inpainting for categorical image data from the National Land Cover Database for producing a diverse set of land use counterfactual scenarios. We find that this approach is effective for producing a distribution of realistic image completions in certain masking configurations. However, the resulting distribution is not well-calibrated in terms of spatial summary statistics commonly used with LULC data and exhibits substantial underdispersion.

Shadow detection and removal is a challenging problem in the analysis of hyperspectral images. Yet, this step is crucial for analyzing data for remote sensing applications like methane detection. In this work, we develop a shadow detection and removal method only based on the spectrum of each pixel and the overall distribution of spectral values. We first introduce Iterative Logistic Regression(ILR) to learn a spectral basis in which shadows can be linearly classified. We then model the joint distribution of the mean radiance and the projection coefficients of the spectra onto the above basis as a parametric linear combination of Gaussians. We can then extract the maximum likelihood mixing parameter of the Gaussians to estimate the shadow coverage and to correct the shadowed spectra. Our correction scheme reduces correction artefacts at shadow borders. The shadow detection and removal method is applied to hyperspectral images from MethaneAIR, a precursor to the satellite MethaneSAT.

We study self-supervised representation learning on high-dimensional data under resource constraints. Our work is motivated by applications of vision transformers to weather and climate data. Such data frequently comes in the form of tensors that are both higher dimensional and of larger size than the RGB imagery one encounters in many computer vision experiments. This raises scaling issues and brings up the need to leverage available compute resources efficiently. Motivated by results on masked autoencoders, we show that it is possible to use sampling of subtensors as the sole augmentation strategy for contrastive learning with a sampling ratio of $\sim$1\%. This is to be compared to typical masking ratios of $75\%$ or $90\%$ for image and video data respectively. In an ablation study, we explore extreme sampling ratios and find comparable skill for ratios as low as $\sim$0.0625\%. Pursuing efficiencies, we are finally investigating whether it is possible to generate robust embeddings on dimension values which were not present at training time. We answer this question to the positive by using learnable position encoders which have continuous dependence on dimension values.

Human health is negatively impacted by poor air quality including increased risk for respiratory and cardiovascular disease. Due to a recent increase in extreme air quality events, both globally and specifically in the United States, finer resolution air quality forecasting guidance is needed to effectively adapt to these events. The National Oceanic and Atmospheric Administration provides air quality forecasting guidance for the United States. Their air quality forecasting model is currently forecasting at a 15 km resolution, however for improved forecast skill, the goal is to reach a 3 km resolution. This is currently not feasible due prohibitive computational needs for the transport of chemical species. In this work we describe a deep learning transport emulator that is able to reduce computations and maintain skill comparable with the existing model. We show how this method performs well in the presence of extreme air quality events, making it a potential candidate for operational use in the near-term future. We also evaluate how well this model maintains the physical properties of the modeled transport.

Extreme weather events in 2019 prevented U.S. farmers from planting crop on a record 19.4 million acres, more than double the previous record. Insurance reports show the majority of these were intended to be corn acres and prevented due to excess soil moisture and precipitation. However, we still lack a detailed understanding of how weather and soil conditions lead to prevented planting, as well as how climate change may impact future outcomes. Machine learning provides a promising approach to this challenge. Here, we model the drivers of prevented corn planting using soil characteristics, monthly weather conditions, and geospatial information. Due to the extreme nature of events causing prevented planting, we use a novel-design zero-inflated regression (ZIR) model that predicts the occurrence of prevented planting as well as the potential severity. We identify key environmental drivers of prevented planting, including May rainfall and soil drainage class. Under climate change scenarios, the model interestingly projects future instances of prevented planting to be less frequent but more severe relative to the historical period.

Cooperative Logistics studies the setting where logistics companies pool their resources together to improve their individual performance. Prior literature suggests carbon savings of approximately 22%. If attained globally, this equates to 440,000,000 tonnes of CO2-eq. Whilst well-studied in operations research –industrial adoption remains limited due to a lack of trustworthy cooperation. A key remaining challenge is fair and scalable gain sharing (i.e., how much should each company be fairly paid?). We propose the use of deep reinforcement learning with a neural reward model for coalition structure generation and present early findings.

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.

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.

Sodium-ion batteries (SIBs) have been gaining attention for applications like grid-scale energy storage, largely owing to the abundance of sodium and an expected favorable $/kWh figure. Improving the performance of SIB electrode materials will enable these batteries to compete with mature technologies like lithium-ion batteries (LIBs) at scale. SIBs can leverage the well-established manufacturing knowledge of LIBs, but several materials synthesis and performance challenges for electrode materials need to be addressed for SIBs to mature to an industrial scale. This work extracts challenges in the performance and synthesis of SIB cathode active materials (CAMs) and reviews corresponding mitigation strategies from a combination of SIB and related LIB literature employing custom natural language processing (NLP) tools. These NLP tools help in identifying the mitigation strategies of interest and subsequently evaluate them using a process-based cost model and other scalability metrics. This approach facilitates the generation of quantitative insights and enables a unique comparison among a broad set of lab-proposed mitigation strategies. These derived insights enable engineers in research and industry to navigate a large number of proposed strategies and focus on impactful scalability-informed mitigation strategies to accelerate the transition from lab to commercialization.

Earth system models (ESMs) are vital for understanding past, present, and future climate, but they suffer from legacy technical infrastructure. ESMs are primarily implemented in Fortran, a language that poses a high barrier of entry for early career scientists and lacks a GPU runtime, which has become essential for continued advancement as GPU power increases and CPU scaling slows. Fortran also lacks differentiability — the capacity to differentiate through numerical code — which enables hybrid models that integrate machine learning methods. Converting an ESM from Fortran to Python/JAX could resolve these issues. This work presents a semi-automated method for translating individual model components from Fortran to Python/JAX using a large language model (GPT-4). By translating the photosynthesis model from the Community Earth System Model (CESM), we demonstrate that the Python/JAX version results in up to 100x faster runtimes using GPU parallelization, and enables parameter estimation via automatic differentiation. The Python code is also easy to read and run and could be used by instructors in the classroom. This work illustrates a path towards the ultimate goal of making climate models fast, inclusive, and differentiable.

Several studies have indicated that delivering insights and feedback on water usage has been effective in curbing water consumption, making it a pivotal component in achieving long-term sustainability objectives. Despite a significant proportion of water consumption originating from large residential and commercial buildings, there is a scarcity of cost-effective and easy-to-integrate solutions that provide water usage insights in such structures. Furthermore, existing methods for disaggregating water usage necessitate training data and rely on frequent data sampling to capture patterns, both of which pose challenges when scaling up and adapting to new environments.In this work, we aim to solve these challenges through a novel end-to-end approach which records data from pressure sensors and uses time-series classification by DNN models to determine room-wise water consumption in a building. This consumption data is then fed to a novel water disaggregation algorithm which can suggest a set of water-usage events, and has a flexible requirement of training data and sampling granularity. We conduct experiments using our approach and demonstrate its potential as a promising avenue for in-depth exploration, offering valuable insights into water usage on a large scale.

Electrifying heavy-duty trucks offers a substantial opportunity to curtail carbon emissions, advancing toward a carbon-neutral future. However, the inherent challenges of limited battery energy and the sheer weight of heavy-duty trucks lead to reduced mileage and prolonged charging durations. Consequently, battery-swapping services emerge as an attractive solution for these trucks. This paper employs a two-fold approach to investigate the potential and enhance the efficacy of such services. Firstly, spatial-temporal demand prediction models are adopted to predict the traffic patterns for the upcoming hours. Subsequently, the prediction guides an optimization module for efficient battery allocation and deployment. Analyzing the heavy-duty truck data on a highway network spanning over 2,500 miles, our model and analysis underscore the value of prediction/machine learning in facilitating future decision-makings. In particular, we find that the initial phase of implementing battery-swapping services favors mobile battery-swapping stations, but as the system matures, fixed-location stations are preferred.

Environment plays an important role in health in/equity and it holds potential for community-level intervention. However, there are important gaps in the literature about the impact of changing environment on an individual's health across countries. The overarching objective of this paper is to examine how changes in the environment (green spaces, temperature, night-time lights, built environment, etc.) influence health equity by applying Multi-dimensional LSTM (M-LSTM) to routine collected data for people living in diverse environments. We developed and validated a data fusion approach to predict malaria incidence rate for the year 2017 using spatio-temporal data from 2000 - 2016 across three South Asian countries: Pakistan, India and Bangladesh. The proposed M-LSTM model improves prediction by 1.75% compared to Conv-LSTM model on all South Asian countries. Additionally, M-LSTM also outperforms Random Forest. The data and code will be made available.

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.

This study presents the application of generative deep learning techniques to evaluate marine fog visibility conditions by nowcasting visibility 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 Vaisala Weather Transmitter model WXT50, mounted on the Research Vessel (R/V) Atlantic Condor. To perform nowcasting, the time series of fog visibility (Vis), wind speed (Uh), dew point depression (Ta-Td), and relative humidity (RHw) 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 cGAN error was not too far off. At both lead times for Vis < 1 km, the cGAN model tracked the variation in Vis very well, which suggests that there is potential for future generative analysis of marine fog formation using observational meteorological parameters.

Identifying stop-and-go events (SAGs) in traffic flow presents an important avenue for advancing data-driven research for climate change mitigation and sustainability, owing to their substantial impact on carbon emissions, travel time, fuel consumption, and roadway safety. In fact, SAGs are estimated to account for 33-50\% of highway driving externalities. However, insufficient attention has been paid to precisely quantifying where, when, and how much these SAGs take place– necessary for downstream decision making, such as intervention design and policy analysis. A key challenge is that the data available to researchers and governments are typically sparse and aggregated to a granularity that obscures SAGs. To overcome such data limitations, this study thus explores the use of traffic reconstruction techniques for SAG identification. In particular, we introduce a kernel-based method for identifying spatio-temporal features in traffic and leverage bootstrapping to quantify the uncertainty of the reconstruction process. Experimental results on California highway data demonstrate the promise of the method for capturing SAGs. This work contributes to a foundation for data-driven decision making to advance sustainability of traffic systems.

Climate change has led to an increased frequency of natural disasters such as floods and cyclones. This emphasizes the importance of effective disaster monitoring. In response, the remote sensing community has explored change detection methods. These methods are primarily categorized into supervised techniques, which yield precise results but come with high labeling costs, and unsupervised techniques, which eliminate the need for labeling but involve intricate hyperparameter tuning. To address these challenges, we propose a novel unsupervised change detection method named Prototype-oriented Unsupervised Change Detection for Disaster Management (PUCD). PUCD captures changes by comparing features from pre-event, post-event, and prototype-oriented change synthesis images via a foundational model, and refines results using the Segment Anything Model (SAM). Although PUCD is an unsupervised change detection, it does not require complex hyperparameter tuning. We evaluate PUCD framework on the LEVIR-Extension dataset and the disaster dataset and it achieves state-of-the-art performance compared to other methods on the LEVIR-Extension dataset.

Over-consumption of services and products leads to potential natural resource exhaustion, high environmental impact, and societal inequalities. Individuals can achieve more sustainable consumption by drastically changing their lifestyle choices and potentially sacrificing personal values or wishes. Conversely, achieving sustainable consumption while accounting for personal values is a more complex challenge, as potential trade-offs arise when trying to satisfy sustainability and personal goals. This article focuses on the value-sensitive design of recommender systems with neural networks and genetic algorithms to support consumers to shop more sustainably while respecting their personal preferences. We formalize recommendations as a multi-objective optimization problem, where each objective represents different sustainability goals and personal values. Recommendations are generated and evaluated on a synthetic historical dataset based on real-world synthetic data sources. The results indicate considerable environmental impact, without extreme personal sacrifices when consumers accept only a fraction of the recommendations.

Healthy urban greenery is a fundamental asset to mitigate climate change phenomenons such as extreme heat and air pollution. However, urban trees are often affected by abiotic and biotic stressors that hamper their functionality, and whenever not timely managed, even their survival. The current visual or instrumented inspection techniques often require a high amount of human labor making frequent assessments infeasible at a city-wide scale. In this work, we present the GreenScan Project, a ground-based sensing system designed to provide health assessment of urban trees at high space-time resolutions, with low costs. The system utilises thermal and multi-spectral imaging sensors, fused using computer vision models to estimate two tree health indexes, namely NDVI and CTD. Preliminary evaluation of the system was performed through data collection experiments in Cambridge, USA. Overall, this work illustrates the potential of autonomous mobile ground-based tree health monitoring on city-wide scales at high temporal resolutions with low-costs.

In remote sensing, Change Detection (CD) refers to locating surface changes in the same area over time. Changes can occur due to man-made or natural activities, and CD is important for analyzing climate changes. The recent advancements in satellite imagery and deep learning allow the development of affordable and powerful CD solutions. The breakthroughs in computer vision Foundation Models (FMs) bring new opportunities for better and more flexible remote sensing solutions. However, solving CD using FMs has not been explored before and this work presents the first FM-based deep learning model, SAM-CD. We propose a novel model that adapts the Segment Anything Model (SAM) for solving CD. The experimental results show that the proposed approach achieves the state of the art when evaluated on two challenging benchmark public datasets LEVIR-CD and DSIFN-CD.

Geographic location is essential for modeling tasks in climate-related fields ranging from ecology to the Earth system sciences. Here, a meaningful feature representation of locations is highly helpful as a description that encodes location-specific aspects. However, obtaining such a representation is challenging and requires an algorithm to distill semantic information of one location from available data. To address this challenge, we introduce GeoCLIP, a global, general-purpose geographic location encoder that provides vector embeddings summarizing the characteristics of a given location for convenient usage in diverse downstream tasks. We show that GeoCLIP embeddings, pretrained on multi-spectral Sentinel-2 satellite data, can be used for various predictive out-of-domain tasks, including temperature prediction and animal recognition in imagery, and outperform existing competing approaches. This demonstrates the potential of general-purpose location encoders and opens the door to learning meaningful representations of our planet from the vast, varied, and largely untapped modalities of geospatial data.

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.

High-resolution simulations, such as the ICOsahedral Non-hydrostatic Large-Eddy Model (ICON-LEM), provide valuable insights into the complex interactions among aerosols, clouds, and precipitation, which are the major contributors to climate change uncertainty. However, due to its exorbitant computational costs, it can only be employed for a limited period and geographical area. To address this, we propose a more cost-effective method powered by emerging machine learning approach -- leveraging high-resolution climate simulation as the oracle and abundant unlabeled data drawn from satellite data -- to better understand the intricate dynamics of the climate system. Our approach involves active learning techniques to predict autoconversion rates, a crucial step in precipitation formation, while significantly reducing the need for a large number of labeled instances. In this study, we present novel methods: custom query strategy fusion for labeling instances, WiFi and MeFi, along with active feature selection based on SHAP, designed to tackle real-world challenges due to its simplicity and practicality in application, specifically focusing on the prediction of autoconversion rates.

High-fidelity topography models of the bedrock underneath the thick Antarctic ice sheet can improve scientists’ understanding of ice flow and its contributions to global sea level rise. However, the bed topography of Antarctica is one of the most challenging surfaces on Earth to map, requiring airplanes with ice-penetrating radars to survey the vast and remote continent. We propose a model that leverages readily available surface topography data from satellites as an auxiliary input modality for bed topography super-resolution. We use a non-parametric Gaussian Process model to transfer local, non-stationary covariance patterns from surface to bedrock. In a controlled reconstruction experiment over complex East Antarctic terrain, our proposed method outperforms bicubic interpolation at all five tested magnification factors, reducing RMSE by 67% at x2 and 25% at x6 magnification. This work demonstrates the opportunity for data fusion methods to advance downstream climate modelling and steward climate change adaptation strategies.

Climate perceptions shared on social media are an invaluable barometer of public attention. By directing research towards this topic, we can eventually improve the effectiveness of climate change communication, increase public engagement, and enhance climate change education. We propose two real-world image datasets to promote impactful research both in the Computer Vision community and beyond. Firstly, ClimateTV, a dataset containing over 700,000 climate change-related images posted on Twitter and labelled on basis of the image hashtags. Secondly, ClimateCT, a Twitter dataset containing images with five-dimensional annotations in super-categories (i) Animals, (ii) Climate action, (iii) Consequences, (iv) Setting, and (v) Type. These challenging classification datasets contain classes which are designed according to their relevance in the context of climate change. The challenging nature of the datasets is given by varying class diversities (e.g. polar bear vs. land mammal) and foci (e.g. arctic vs. snowy residential area). The analyses of our datasets using CLIP embeddings and query optimization (CoCoOp) further showcase the challenging nature of ClimateTV and ClimateCT.

Climate change and other anthropogenic factors have led to a catastrophic decline in insects, endangering both biodiversity and the ecosystem services on which human society depends. Data on insect abundance, however, remains woefully inadequate. Camera traps, conventionally used for monitoring terrestrial vertebrates, are now being modified for insects, especially moths. We describe a complete, open-source machine learning-based software pipeline for automated monitoring of moths via camera traps, including object detection, moth/non-moth classification, fine-grained identification of moth species, and tracking individuals. We believe that our tools, which are already in use across three continents, represent the future of massively scalable data collection in entomology.

We introduce a novel Graph Attention Autoencoder (GAE) with spatial regularization to address the challenge of scalable anomaly detection in spatiotemporal rainfall data across India from 1990 to 2015. Our model leverages a Graph Attention Network (GAT) to capture spatial dependencies and temporal dynamics in the data, further enhanced by a spatial regularization term ensuring geographic coherence. We construct two graph datasets employing rainfall, pressure, and temperature attributes from the Indian Meteorological Department and ERA5 Reanalysis on Single Levels, respectively. Our network operates on graph representations of the data, where nodes represent geographic locations, and edges, inferred through event synchronization, denote significant co-occurrences of rainfall events. Through extensive experiments, we demonstrate that our GAE effectively identifies anomalous rainfall patterns across the Indian landscape. Our work paves the way for sophisticated spatiotemporal anomaly detection methodologies in climate science, contributing to better climate change preparedness and response strategies.

Fossil fuel energy production is a leading cause of climate change. While wind and solar energy have made advancements, ocean waves, a more consistent clean energy source, remain underutilized. Wave Energy Converters (WEC) transform wave power into electric energy. To be economically viable, modern WECs need sophisticated real-time controllers that boost energy output and minimize mechanical stress, thus lowering the overall cost of energy (LCOE). This paper presents how a Reinforcement Learning (RL) controller can outperform the default spring damper controller for complex spread waves in the sea, enhancing wave energy's viability. Using the Proximal Policy Optimization (PPO) algorithm with Transformer variants as function approximators, the RL controllers optimize multi-generator Wave Energy Converters (WEC), leveraging wave sensor data for multiple cost-efficiency goals. After successful tests in the EuropeWave\footnote{EuropeWave: https://www.europewave.eu/} project's emulator tank, the platform is planned to deploy. We discuss the challenges of deployment at the BiMEP site and how we had to tune the RL controller to address that. The RL controller outperforms the default Spring Damper controller in the BiMEP\footnote{BiMEP: https://www.bimep.com/en/} conditions by 22.8% on energy capture. Enhancing wave energy's economic viability will expedite the transition to clean energy, reducing carbon emissions and fostering a healthier climate.

Climate and poverty are intrinsically related: regions with extreme temperatures, large temperature variability, and recurring extreme weather events tend to be ranked among the poorest and most vulnerable to climate change. Nevertheless, there currently is no established method to directly estimate the impact of specific climate variables on poverty and to identify geographical regions at high risk of being negatively affected by climate change. In this work, we propose a new approach based on Graph Neural Networks (GNNs) to estimate the effect of climate and remote sensing variables on poverty indicators measuring Education, Health, Living Standards, and Income. Furthermore, we use the trained models and perturbation analyses to identify the geographical regions most vulnerable to the potential variations in climate variables.

With the advances in large scale Foundation Models, web scale access to sustainability related data, planetary scale satellite data, the opportunity for community to collaborate with AI tools to create positive impact regarding climate change is imminent. Hence, the need for development of AI co-pilots in the field. In this work, we start with design of early stage Sustainability AI copilot (SAI). We demonstrate an early prototype of SAI with 4 use cases. Given the carbon footprint of industry, SAI enables sustainability officers at companies to understand regulatory frameworks, and discover sustainability best practices from peers to plan corporate sustainability roadmap. SAI enables climate enthusiasts to protect vulnerable people from climate disasters by exploring planetary scale satellite data. SAI also ideates with visionaries to explore climate friendly product designs. SAI also enables activists to create awareness about inclusion of vulnerable and persons with disability in the climate conversation. In the era of large scale AI, this work proposes the opportunity for creating co-pilots to accelerate progress on Sustainable Development Goals

Cloud analysis is a critical component of weather and climate science, impacting various sectors like disaster management. However, achieving fine-grained cloud analysis, such as cloud segmentation, in remote sensing remains challenging due to the inherent difficulties in obtaining accurate labels, leading to significant labeling errors in training data. Existing methods often assume the availability of reliable segmentation annotations, limiting their overall performance. To address this inherent limitation, we introduce an innovative model-agnostic Cloud Adaptive-Labeling (CAL) approach, which operates iteratively to enhance the quality of training data annotations and consequently improve the performance of the learned model. Our methodology commences by training a cloud segmentation model using the original annotations. Subsequently, it introduces a trainable pixel intensity threshold for adaptively labeling the cloud training images on-the-fly. The newly generated labels are then employed to fine-tune the model. Extensive experiments conducted on multiple standard cloud segmentation benchmarks demonstrate the effectiveness of our approach in significantly boosting the performance of existing segmentation models. Our CAL method establishes new state-of-the-art results when compared to a wide array of existing alternatives.

Climate change is inducing significant range shifts in insects and other organisms. Large-scale temporal data on populations and distributions are essential for quantifying the effects of climate change on biodiversity and ecosystem services, providing valuable insights for both conservation and pest management. With images from camera traps, we aim to use Mahalanobis distance-based confidence scores to automatically detect new moth species in a region. We intend to make out-of-distribution detection interpretable by identifying morphological characteristics of different species using Grad-CAM. We hope this algorithm will be a useful tool for entomologists to study range shifts and inform climate change adaptation.

Pesticides, widely used in agriculture for pest control, contribute to the climate crisis. Integrated pest management (IPM) is preferred as a climate-smart alternative. However, low adoption rates of IPM are observed due to farmers' skepticism about its effectiveness, so we introduce an enhancing data analysis framework for IPM to combat that. Our framework provides i) robust pest population predictions across diverse environments with invariant and causal learning, ii) interpretable pest presence predictions using transparent models, iii) actionable advice through counterfactual explanations for in-season IPM interventions, iv) field-specific treatment effect estimations, and v) causal inference to assess advice effectiveness.

Sustainable global development is one of the most prevalent challenges facing the world today, hinging on the equilibrium between socioeconomic growth and environmental sustainability. We propose approaches to monitor and quantify sustainable development along the Shared Socioeconomic Pathways (SSPs), including mathematically derived scoring algorithms, and machine learning methods. These integrate socioeconomic and environmental datasets, to produce an interpretable metric for SSP alignment. An initial study demonstrates promising results, laying the groundwork for the application of different methods to the monitoring of sustainable global development.

Effective action is crucial in order to avert climate disaster. Key in enacting change is the swift adoption of climate positive legislation which advocates for climate change mitigation and adaptation. This is because government legislation can result in far-reaching impact, due to the relationships between climate policy, technology, and market forces. To advocate for legislation, current strategies aim to identify potential levers and obstacles, presenting an opportunity for the application of recent advances in machine learning language models. Here we propose a machine learning pipeline to analyse climate legislation, aiming to investigate the feasibility of natural language processing for the classification of climate legislation texts, to predict policy voting outcomes. By providing a model of the decision making process, the proposed pipeline can enhance transparency and aid policy advocates and decision makers in understanding legislative decisions, thereby providing a tool to monitor and understand legislative decisions towards climate positive impact.

This proposal advocates a comprehensive and systematic analysis aimed at mapping and characterizing the intricate landscape of Artificial Intelligence and Machine Learning applications and their impacts within the domain of climate change research, both in adaption and mitigation efforts. Notably, a significant upswing in this interdisciplinary intersection has been observed since 2020. Utilizing advanced topic clustering techniques and qualitative analysis, we have discerned 12 distinct macro areas that supplement, enrich, and expand upon those identified in prior research. The primary objective of this undertaking is to furnish a data-rich panoramic view and informative insights regarding the functions and tools of the mentioned disciplines. Our intention is to offer valuable guidance to the scholarly community and propel further research endeavors, encouraging meticulous examinations of research trends and gaps in addressing the formidable challenges posed by climate change and the climate crisis.

This study addresses the accuracy challenge in Global Land Use and Land Cover (LULC) maps, crucial for policy making towards climate change mitigation. We evaluate two LULC products based on advanced machine learning techniques across two representative nations, Ecuador and Germany, employing a novel accuracy metric. The analysis unveils a notable accuracy enhancement in the convolutional neural network-based approach against the random forest model used for comparison. Our findings emphasize the potential of sophisticated machine learning methodologies in advancing LULC mapping accuracy, an essential stride towards data-driven, climate-relevant land management and policy decisions.

As the world moves towards a clean energy future to mitigate the risks of climate change, the discovery of new catalyst materials plays a significant role in enabling the sustainable production and transformation of energy [2]. The development and verification of fast, accurate, and efficient artificial intelligence and machine learning techniques is critical to shortening time-intensive calculations, reducing costs, and improving computational feasibility. We propose applying the Crystal Hamiltonian Graph Neural Network (CHGNet) on the OC20 dataset in order to iteratively perform structure-to-energy and forces calculations and identify the lowest energy across relaxed structures for a given adsorbate-surface combination. CHGNet's predictions will be compared and benchmarked to corresponding values calculated by density functional theory (DFT) [7] and other models to determine its efficacy.

Electrification has emerged as a pivotal trend in the energy transition to address climate change, leading to a substantial surge in the demand for batteries. Accurately predicting the internal states and performance of batteries assumes paramount significance, as it ensures the safe and stable operation of batteries and informs decision-making processes, such as optimizing battery operation for arbitrage opportunities. However, current models struggle to strike a balance between precision and computational efficiency or are limited in their applicability to specific scenarios. We aim to adopt a physics-informed deep operator network (PI-DeepONet) for internal battery state estimation based on the rigorous P2D model, which can simultaneously achieve high precision and computational efficiency. Furthermore, it exhibits promising prospects for extension beyond lithium-ion batteries to encompass various battery technologies.

The maritime industry faces an unprecedented challenge in the form of decarbonization. With strict emissions reduction targets in place, the industry is turning to machine learning-based decision support models to achieve sustainability goals. This proposal explores the transformative potential of digitalization and machine learning approaches in maritime operations, from optimizing ship speeds to enhancing supply chain management. By examining various machine learning techniques, this work provides a roadmap for reducing emissions while improving operational efficiency in the maritime sector.

Globally, buildings account for 30% of end-use energy consumption and 27% of energy sector emissions, and yet the building sector is lacking in low-temporal-latency, high-spatial-resolution data on energy consumption and resulting emissions. Existing methods tend to either have low resolution, high latency (often a year or more), or rely on data typically unavailable at scale (such as self-reported energy consumption). We propose a machine learning based bottom-up model that combines satellite-imagery-derived features to compute Scope 1 global emissions estimates both for residential and commercial buildings at a 1 square km resolution with monthly global updates.

As the major ingredient of concrete and asphalt, sand is vital to economic growth, and will play a key role in aiding the transition to a low carbon society. However, excessive and unregulated sand mining in the Global South has high socio-economic and environmental costs, and amplifies the effects of climate change. Sand mines are characterized by informality and high temporal variability, and data on the location and extent of these mines tends to be sparse. We propose to build custom sand-mine detection tools by fine-tuning foundation models for earth observation, which leverage self supervised learning - a cost-effective and powerful approach in sparse data regimes. These tools allow for real-time monitoring of sand mining activity and can enable more effective policy and regulation.

A significant fraction (7-15%) of renewable energy generated goes into waste in the grids around the world today due to oversupply issues and transmission constraints. Being able to predict when and where renewable curtailment occurs would improve renewable utilization. The core of this work is to enable the machine learning community to help decarbonize electricity grids by unlocking the potential of renewable energy through curtailment prediction.

Securing sufficient forecast lead time for local precipitation is essential for preventing hazardous weather events.Nonetheless, global warming-induced climate change is adding to the challenge of accurately predicting severe precipitation events, such as heavy rainfall.In this work, we propose a deep learning-based precipitation post-processor approach to numerical weather prediction (NWP) models.The precipitation post-processor consists of (i) self-supervised pre-training, where parameters of encoder are pre-trained on the reconstruction of masked variables of the atmospheric physics domain, and (ii) transfer learning on precipitation segmentation tasks (target domain) from the pre-trained encoder.We also introduce a heuristic labeling approach for effectively training class-imbalanced datasets.Our experiment results in precipitation correction for regional NWP show that the proposed method outperforms other approaches.

We introduce EarthPT -- an Earth Observation (EO) pretrained transformer. EarthPT is a 700 million parameter decoding transformer foundation model trained in an autoregressive self-supervised manner and developed specifically with EO use-cases in mind. We demonstrate that EarthPT is an effective forecaster that can accurately predict future pixel-level surface reflectances across the 400-2300 nm range well into the future. For example, forecasts of the evolution of the Normalised Difference Vegetation Index (NDVI) have a typical error of approximately 0.05 (over a natural range of -1 -> 1) at the pixel level over a five month test set horizon, out-performing simple phase-folded models based on historical averaging. We also demonstrate that embeddings learnt by EarthPT hold semantically meaningful information and could be exploited for downstream tasks such as highly granular, dynamic land use classification. Excitingly, we note that the abundance of EO data provides us with -- in theory -- quadrillions of training tokens. Therefore, if we assume that EarthPT follows neural scaling laws akin to those derived for Large Language Models (LLMs), there is currently no data-imposed limit to scaling EarthPT and other similar ‘Large Observation Models.’

Mapping aquaculture ponds is critical for restoration, conservation, and climate adaptation efforts. Aquaculture can contribute to high levels of water pollution from untreated effluent and negatively impact coastal ecosystems. Large-scale aquaculture is also a significant driver in mangrove deforestation, thus reducing the world’s carbon sinks and exacerbating the effects of climate change. However, finding and mapping these ponds on the ground can be highly labor and time-intensive. Most existing automated techniques are focused only on spatial location and do not consider production intensification, which is also crucial to understanding their impact on the surrounding ecosystem. We can classify them into two main types: a) Extensive ponds, which are large, irregularly-shaped ponds that rely on natural productivity, and b) intensive ponds which are smaller and regularly shaped. Intensive ponds use machinery such as aerators that maximize production and also result in the characteristic presence of air bubbles on the pond’s surface. The features of these two types of ponds make them distinguishable and detectable from satellite imagery.In this tutorial, we will discuss types of aquaculture ponds in detail and demonstrate how they can be detected and classified using satellite imagery. The tutorial will introduce an open dataset of human-labeled aquaculture ponds in the Philippines and Indonesia. Using this dataset, the tutorial will use semantic segmentation to map out similar ponds over an entire country and classify them as either extensive or intensive, going through the entire process of i) satellite imagery retrieval, ii) preprocessing these images into a training-ready dataset, iii) model training, and iv) finally model rollout on a sample area. Throughout, the tutorial will leverage PyTorch Lightning, a machine learning framework that provides a simplified and streamlined interface for model experimentation and deployment. This tutorial aims to discuss the relevance of aquaculture ponds in climate adaptation and equip users with the necessary inputs and tools to perform their own ML-powered earth observation projects.

Transportation contributes to 29% of all greenhouse gas (GHG) emissions in the US, of which 58% are from light-duty vehicles and 28% from medium-to-heavy duty vehicles (MHDVs) [1]. Battery electric vehicles (EVs) emit 90% less life cycle GHGs than their internal combustion engine (ICEV) counterparts [2], but currently only comprise 2% of all vehicles in the U.S. EVs thus represent a crucial step in decarbonizing road transportation. One major challenge in replacing ICEVs with EVs at scale is the ability to charge a large number of EVs within the constraints of power systems in a cost-effective way. This is an especially prominent problem for MHDVs used in commercial fleets such as shuttle buses and delivery trucks, as they generally require more energy to complete assigned trips compared to light-duty vehicles. In this tutorial, we describe the myriad challenges in charging EVs at scale and define common objectives such as minimizing total load on power systems, minimizing fleet operating costs, as well as maximizing vehicle state of charge and onsite photovoltaic energy usage. We discuss common constraints such as vehicle trip energy requirements, charging station power limits, and limits on vehicles’ time to charge between trips. We survey several different methods to formulate EV charging and energy dispatch as a mathematically solvable optimization problem, using tools such as convex optimization, Markov decision process (MDP), and reinforcement learning (RL). We introduce a commercial application of model-based predictive control (MPC) algorithm, ZEVi (Zero Emission Vehicle intelligence), which solves optimal energy dispatch strategies for charging sessions of commercial EV fleets. Using a synthetic dataset modeled after a real fleet of electric school buses, we engage the audience with a hands-on exercise applying ZEVi to find the optimal charging strategy for a commercial fleet. Lastly, we briefly discuss other contexts in which methods originating from process control and deep learning, like MPC and RL, can be applied to solve problems related to climate change mitigation and adaptation. With the examples provided in this tutorial, we hope to inspire the audience to come up with their own creative ways to apply these methods in different fields within the climate domain. References[1] EPA (2023). Inventory of U.S. Greenhouse Gas Emissions and Sinks: 1990-2021. U.S. Environmental Protection Agency, EPA 430-R-23-002. [2] Verma, S., Dwivedi, G., & Verma, P. (2022). Life cycle assessment of electric vehicles in comparison to combustion engine vehicles: A review. Materials Today: Proceedings, 49, 217-222.

Machine learning models for parsing remote sensing data have a wide range of societally relevant applications, but labels used to train these models can be difficult or impossible to acquire. This challenge has spurred research into self-supervised learning for remote sensing data. Current self-supervised learning approaches for remote sensing data draw significant inspiration from techniques applied to natural images. However, remote sensing data has important differences from natural images -- for example, the temporal dimension is critical for many tasks and data is collected from many complementary sensors. We show we can create significantly smaller performant models by designing architectures and self-supervised training techniques specifically for remote sensing data. We introduce the Pretrained Remote Sensing Transformer (Presto), a transformer-based model pre-trained on remote sensing pixel-timeseries data. Presto excels at a wide variety of globally distributed remote sensing tasks and performs competitively with much larger models while requiring far less compute. Presto can be used for transfer learning or as a feature extractor for simple models, enabling efficient deployment at scale.

We propose a framework for applying reinforcement learning to contextual two-stage stochastic optimization and apply this framework to the problem of energy market bidding of an off-shore wind farm. Reinforcement learning could potentially be used to learn close to optimal solutions for first stage variables of a two-stage stochastic program under different contexts. Under the proposed framework, these solutions would be learned without having to solve the full two-stage stochastic program. We present initial results of training using the DDPG algorithm and present intended future steps to improve performance.

Clouds containing ice particles play a crucial role in the climate system. Yet they remain a source of great uncertainty in climate models and future climate projections. In this work, we create a new observational constraint of regime-dependent ice microphysical properties at the spatio-temporal coverage of geostationary satellite instruments and the quality of active satellite retrievals. We achieve this by training a convolutional neural network on three years of SEVIRI and DARDAR data sets. This work will enable novel research to improve ice cloud process understanding and hence, reduce uncertainties in a changing climate and help assess geoengineering methods for cirrus clouds.

Satellite-based remote sensing is instrumental in the monitoring and mitigation of the effects of anthropogenic climate change. Large scale, high resolution data derived from these sensors can be used to inform intervention and policy decision making, but the timeliness and accuracy of these interventions is limited by use of optical data, which cannot operate at night and is affected by adverse weather conditions. Synthetic Aperture Radar (SAR) offers a robust alternative to optical data, but its associated complexities limit the scope of labelled data generation for traditional deep learning. In this work, we apply a self-supervised pretraining scheme, masked autoencoding, to SAR amplitude data covering 8.7\% of the Earth's land surface area, and tune the pretrained weights on two downstream tasks crucial to monitoring climate change - vegetation cover prediction and land cover classification. We show that the use of this pretraining scheme reduces labelling requirements for the downstream tasks by more than an order of magnitude, and that this pretraining generalises geographically, with the performance gain increasing when tuned downstream on regions outside the pretraining set. Our findings significantly advance climate change mitigation by facilitating the development of task and region-specific SAR models, allowing local communities and organizations to deploy tailored solutions for rapid, accurate monitoring of climate change effects.

The growing penetration of intermittent, renewable generation in US power grids results in increased operational uncertainty. In that context, accurate forecasts are critical, especially for wind generation, which exhibits large variability and is historically harder to predict. To overcome this challenge, this work proposes a novel Bundle-Predict-Reconcile (BPR) framework that integrates asset bundling, machine learning, and forecast reconciliation techniques to accurately predict wind power at the asset, bundle, and fleet level. Notably, our approach effectively introduces an auxiliary learning task (predicting the bundle-level time series) to help the main learning tasks (fleet-level time series) and proposes new asset-bundling criteria to capture the spatio-temporal dynamics of wind power time series. Extensive numerical experiments are conducted on an industry-size dataset of wind farms, demonstrating the benefits of BPR, which consistently and significantly improves forecast accuracy over the baseline approach, especially at the fleet level.

Understanding why travel behavior differs between residents of urban centers and suburbs is key to sustainable urban planning. Especially in light of rapid urban growth, identifying housing locations that minimize travel demand and induced emissions is crucial to mitigate climate change. While the built environment plays an important role, the precise impact on travel behavior is obfuscated by residential self-selection.To address this issue, we propose a double machine learning approach to obtain unbiased, spatially-explicit estimates of the effect of the built environment on travel-related emissions for each neighborhood by controlling for residential self-selection. We examine how socio-demographics and travel-related attitudes moderate the effect and how it decomposes across the 5Ds of the built environment. Based on a case study for Berlin and the travel diaries of 32,000 residents, we find that the built environment causes household travel-related emissions to differ by a factor of almost two between central and suburban neighborhoods in Berlin. To highlight the practical importance for urban climate mitigation, we evaluate current plans for 64,000 new residential units in terms of total induced transport emissions. Our findings underscore the significance of compact development to decarbonize the transport sector.

How areas of land are allocated for different uses, such as forests, urban, and agriculture, has a large effect on carbon balance, and therefore climate change. Based on available historical data on changes in land use and a simulation of carbon emissions/absorption, a surrogate model can be learned that makes it possible to evaluate the different options available to decision-makers efficiently. An evolutionary search process can then be used to discover effective land-use policies for specific locations. Such a system was built on the Project Resilience platform and evaluated with the Land Use Harmonization dataset and the BLUE simulator. It generates Pareto fronts that trade off carbon impact and amount of change customized to different locations, thus providing a potentially useful tool for land-use planning.

The increasing size and severity of wildfires across western North America have generated dangerous concentrations 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 spatial-temporal graph neural network-based PM2.5 forecasting. The experiments in this work focus on determining the optimal time for implementing prescribed fires in California as well as on quantifying the potential trade-offs involved in conducting more prescribed fires outside the fire season.

Data-driven weather prediction (DDWP) models are increasingly becoming popular for weather forecasting. While DDWPs primarily forecast prognostic parameters, many diagnostic meteorological parameters (such as precipitation) are dependent on the most recent weather state and are modeled by learning a data-driven functional mapping of the current meteorological state (c.f. FourCastNet). However, the cost of training bespoke models for diagnostic variables can scale significantly and further limit the use during operationalizing these forecasts. This presents an opportunity to learn dense representations of essential meteorological parameters in a latent space, and using learned representations to model diagnostic parameters, or any other dependent variables. Using learned representations of weather allows for efficient prediction of dependent variables, while dramatically lowering the training cost for such models as well. In this paper, we present one such weather embedding model, WeatherX, trained on decades of reanalysis data that is used to train multiple diagnostic variables. The results indicate that models trained using learned representations of weather offer performance comparable to bespoke models, while leading to significant reduction in resource utilization during training and inference. Further lower memory footprint during operationalization leads to additional gain of running larger ensembles during inference thereby further improving uncertainty quantification of the said forecasts.

Monitoring organic matter is pivotal for maintaining soil health and can help inform sustainable soil management practices. While sensor-based soil information offers higher-fidelity and reliable insights into organic matter changes, sampling and measuring sensor data is cost-prohibitive. We propose a multi-modal, scalable framework that can estimate organic matter from remote sensing data, a more readily available data source while leveraging sparse soil information for improving generalization. Using the sensor data, we preserve underlying causal relations among sensor attributes and organic matter. Simultaneously we leverage inherent structure in the data and train the model to discriminate among domains using contrastive learning. This causal and contrastive constraint minimization ensures improved generalization and adaptation to other domains. We also shed light on the interpretability of the framework by identifying attributes that are important for improving generalization. Identifying these key soil attributes that affect organic matter will aid in efforts to standardize data collection efforts.

As machine learning workloads significantly increase energy consumption, sustainable data centers with low carbon emissions are becoming a top priority for governments and corporations worldwide. There is a pressing need to optimize energy usage in these centers, especially considering factors like cooling, load flexibility based on renewable energy availability, and battery storage utilization. The challenge arises due to the interdependencies of these strategies with fluctuating external factors such as weather and grid carbon intensity. Although there's currently no real-time solution that addresses all these aspects, our proposed Data Center Carbon Footprint Reduction (DCCFR) framework, based on multi-agent Reinforcement Learning (MARL), targets carbon footprint reduction, energy conservation, and cost. Our findings reveal that DCCFR's MARL agents efficiently navigate these complexities, optimizing energy in real-time. Compared to the industry standard ASHRAE controller controlling HVAC for a year in various regions, DCCFR reduced carbon emissions, energy consumption, and energy costs by over 10% with EnergyPlus simulation.

Carbon capture technologies is an important tool for mitigating climate change. In recent years, polymer membrane separation methods have emerged as a promising technology for separating CO2 and other green house gases from the atmosphere. Designing new polymers for such tasks is quite difficult. In this work we look at machine learning based methods to search for new polymer designs optimized for CO2 separation. An ensemble ML models is trained on a large database of molecules to predict permeabilities of CO2/N2 and CO2/O2 pairs. We then use search based optimization to discover new polymers that surpass existing polymer designs. Simulations are then done to verify the predicted performance of the new designs. Overall result suggests that ML based search can be used to discover new polymers optimized for carbon capture.

High-resolution simulations such as the ICOsahedral Non-hydrostatic Large-Eddy Model (ICON-LEM) can be used to study the interactions between aerosols, clouds, and precipitation processes that currently represent the largest source of uncertainty involved in climate change projections. However, due to significant computational costs, it can only be employed for a limited period and area. While machine learning mitigates this, model uncertainties may affect reliability. To address this, we developed a neural network (NN) model powered with evidential learning to assess the data and model uncertainties. Our study focuses on estimating the rate at which small droplets (cloud droplets) collide and coalesce to become larger droplets (raindrops) – autoconversion rates -- as the key process in the precipitation formation, crucial to better understanding cloud responses to anthropogenic aerosols. The results show that the model performs reasonably well, with the inclusion of both aleatoric and epistemic uncertainty estimation, which improves the credibility of the model and provides useful insights for future improvement.

Existing ML-based atmospheric models are not suitable for climate prediction, which requires long-term stability and physical consistency. We present ACE (AI Climate Emulator), a 200M-parameter, autoregressive machine learning emulator of an existing comprehensive 100-km resolution global atmospheric model. The formulation of ACE allows evaluation of physical laws such as the conservation of mass and moisture. The emulator is stable for 10 years, nearly conserves column moisture without explicit constraints and faithfully reproduces the reference model's climate, outperforming a challenging baseline on over 80% of tracked variables. ACE requires nearly 100x less wall clock time and is 100x more energy efficient than the reference model using typically available resources.