Source-linked AI summary
Tackling Climate Change with Machine Learning
David Rolnick, Priya L. Donti, Lynn H. Kaack, Kelly Kochanski, Alexandre Lacoste, Kris Sankaran, Andrew Slavin Ross, Nikola Milojevic-Dupont, Natasha Jaques, Anna Waldman-Brown, Alexandra Luccioni, Tegan Maharaj, Evan D. Sherwin, S. Karthik Mukkavilli, Konrad P. Kording, Carla Gomes, Andrew Y. Ng, Demis Hassabis, John C. Platt, Felix Creutzig, Jennifer Chayes, Yoshua Bengio
TL;DR
Climate change demands coordinated mitigation and adaptation, yet machine-learning practitioners and other fields need clearer guidance on where ML can have high impact. This paper surveys applications across climate and energy domains, drawing on cross-disciplinary expertise and identifying research and deployment opportunities. It emphasizes promising directions while acknowledging subjective evaluation, incomplete coverage, and the limits of technology alone.
Problem
The paper addresses the need to identify where machine learning can best support climate-change mitigation and adaptation amid uncertainty among practitioners and demand from other fields.
Method
The authors provide a cross-domain overview of high-impact ML applications, informed by experts across fields and organized with labels describing leverage, timing, and impact uncertainty.
Results
The paper identifies machine-learning opportunities spanning climate-model improvement, local risk forecasting, electricity systems, and other mitigation and adaptation domains.
Takeaways & Limitations
The recommendations offer researchers, engineers, entrepreneurs, investors, corporate leaders, and governments concrete climate-related problems where ML research or deployment may contribute.
Takeaways & Limitations
The authors state that their flags reflect subjective evaluation, that the paper cannot cover every application, and that technology alone is insufficient.
Abstract
from arXiv · showhide
Climate change is one of the greatest challenges facing humanity, and we, as machine learning experts, may wonder how we can help. Here we describe how machine learning can be a powerful tool in reducing greenhouse gas emissions and helping society adapt to a changing climate. From smart grids to disaster management, we identify high impact problems where existing gaps can be filled by machine learning, in collaboration with other fields. Our recommendations encompass exciting research questions as well as promising business opportunities. We call on the machine learning community to join the global effort against climate change.
Introduction
The paper frames climate change as an urgent, multifaceted mitigation and adaptation challenge and surveys high-impact ways machine learning can help across domains. It combines research opportunities, deployment prospects, and cross-disciplinary recommendations for practitioners and decision-makers.
- Motivation: Climate change is producing increasingly severe storms, droughts, fires, and floods while greenhouse-gas emissions continue to rise.
- Motivation: Mitigation spans electricity, transportation, buildings, industry, and land use, while adaptation requires resilience planning and disaster management.
- Scope and contribution: The paper identifies high-impact applications of machine learning for climate mitigation, adaptation, and tools that enable other strategies.
- Audiences: The recommendations target researchers, entrepreneurs, corporate leaders, and governments, covering research innovation, deployment, efficiency, public services, and planning.
- How to read this paper: The paper flags selected strategies as high leverage, long-term, or uncertain impact, while emphasizing that these labels reflect subjective evaluation and are not definitive.
- Scope and limitations: The authors acknowledge incomplete coverage, the need for collaboration across fields, and that machine learning and technology alone cannot solve climate change.
1 Electricity Systems
Electricity systems account for about a quarter of human-caused greenhouse-gas emissions and face growing demand as other sectors electrify. Machine learning can support the transition, operation, and monitoring of cleaner electricity systems.
- Challenge: About a quarter of human-caused greenhouse-gas emissions come from electricity systems, while electrification will increase demand for low-carbon electricity.
- Challenge: Reducing electricity-system emissions requires rapidly expanding low-carbon sources and phasing out carbon-emitting sources.
- Challenge: Existing CO2-emitting power plants must also reduce emissions because the transition to low-carbon power will not happen overnight.
- Challenge: Because electricity systems exist across countries and contexts, emissions-reduction changes must be implemented broadly.
- Machine-learning opportunities: Machine learning can inform electricity-technology research, deployment, and operation through forecasting, optimization, management, and system monitoring.
1.1 Enabling low-carbon electricity
Enabling low-carbon electricity requires managing variable generation, forecasting supply and demand, optimizing system operations, and accelerating materials discovery. The paper presents machine learning as a tool for these tasks, with domain knowledge and collaboration needed for effective deployment.
- Variable sources: Variable solar and wind generation require controllable sources or storage to buffer changes in output and maintain electricity balance.
- Forecasting supply and demand: Better short-term supply and demand forecasts can reduce reliance on polluting standby plants, while longer-term forecasts support system planning and investment.
- Forecasting supply and demand: Future forecasting models should incorporate weather, climate, physical-system knowledge, and operational objectives such as scheduling costs or greenhouse-gas emissions.
- System optimization: Scheduling and dispatch are slow, complex, and governed by NP-hard optimization problems across timescales from sub-second operation to days ahead.
- System optimization: Machine learning can speed power-system optimization, approximate existing problems, learn control actions, and support dynamic or decentralized scheduling.
- Materials discovery: Machine learning methods can support materials synthesis, characterization, modeling, and design, potentially accelerating technologies for storing or harnessing variable energy.
1.2 Reducing current-system impacts
While electricity systems transition toward low-carbon sources, machine learning can reduce emissions from existing infrastructure by addressing fossil-fuel leakage, delivery losses, and demand management. These applications require careful analysis of their actual greenhouse-gas effects.
- Mitigation opportunities: Current-system mitigation can target fossil-fuel emissions, electricity-delivery waste, and demand flexibility to reduce emissions impacts.
- Fossil-fuel emissions: Machine learning can help detect and prevent methane leakage from natural-gas pipelines and compressor stations using sensor or satellite data.
- Fossil-fuel emissions: Other applications include reducing freight emissions, managing CO2 storage sites, and optimizing power-plant parameters.
- Grid losses: Predictive maintenance can identify proactive electricity-grid upgrades that prevent avoidable resistive losses and associated emissions.
- Demand management: Marginal emissions factors quantify the emissions effects of small demand changes and can support real-time consumer information and multi-day forecasting.
1.3 Ensuring global impact
Machine learning can broaden climate impact beyond data-rich electricity systems by improving clean-energy access and transferring or generating information in low-data settings.
- Approaching low-data settings: Transfer learning can translate insights from high-data to low-data electricity settings because electric grids share underlying system physics.
- Clean electricity access can reduce emissions while supporting social and economic development.It can displace diesel generators, wood-burning stoves, and other carbon-emitting energy sources.
- Satellite imagery combined with image processing, clustering, and optimization can inform electrification initiatives at scale.
- Forecasting demand and power production can help operate rural microgrids, which are harder to balance than country-scale grids.
- Approaching low-data settings: ML can generate missing electricity-system information using satellite imagery, graph search, and cellular-network data.
1.4 Discussion
Electricity-system ML applications can have climate benefits but also carry domain-specific risks, requiring collaboration with decision-makers and practitioners from other fields.
- Reducing emissions in oil and gas through ML could make emissions cheaper and increase them instead.
2 Transportation
Transportation is a major climate challenge where ML can support planning, operations, engineering, electrification, and integration of lower-carbon modes. Some applications have uncertain emissions effects and require incentives aligned with decarbonization.
- Modeling demand: Demand modeling and infrastructure planning can shape trip lengths and transport-mode choices in ways that reduce GHG emissions.
- Modeling demand: ML can estimate origin-destination demand, forecast traffic and transit ridership, and infer travel behavior from novel data.
- Modeling demand: Predicting runway demand and aircraft taxi time can reduce excess fuel burned because of airport congestion.
- Uncertain-impact technologies: Shared mobility and autonomous vehicles have uncertain emissions effects because they may increase vehicle travel and offset efficiency gains.
- Freight routing and consolidation: Freight consolidation and routing can reduce trips, empty returns, and enable rail or water transport for part of a journey.
- Electric vehicles: Electric vehicles are a primary decarbonization pathway, with ML supporting charging, congestion management, vehicle-to-grid algorithms, battery management, and fault detection.
- Electric vehicles: Battery ML models often rely on laboratory data that omit real-world environmental factors, while industry has operational data that could complement academic work.
- Enabling low-carbon options: ML can help integrate transport modes and improve public transit reliability, potentially supporting greater use of low-carbon modes such as rail.
3 Buildings & Cities
Buildings and cities present major opportunities to reduce greenhouse gas emissions, but effective strategies must account for heterogeneous buildings, long lifespans, coordination needs, and uneven data coverage. Machine learning can support tailored building management, urban planning, infrastructure coordination, and low-carbon transitions.
- Buildings: Buildings account for a quarter of global energy-related emissions, while combined efficiency measures could reduce existing-building emissions by up to 90%.The paper notes that some buildings can already consume almost no energy and that many efficiency measures save money.
- Buildings: Building heterogeneity and long lifespans require context-specific strategies for both new construction and retrofits.Optimal interventions vary with age, construction, usage, ownership, and access to low-carbon electricity.
- Building management: Machine learning can model building energy data, optimize smart-building energy use, and tailor interventions to individual buildings.The paper describes applications spanning energy-demand forecasting, energy disaggregation, and intelligent control systems.
- Caveats: Smart-building technologies can create rebound, energy-use, material, privacy, and security concerns, requiring complementary public policies.Reported rebound effects can add 10–20% to building energy consumption in some cases, and cost optimization may not reduce emissions.
- Cities: Urban-scale mitigation requires coordination across districts and cities, where ML can inform policy and optimize heating, cooling, solar generation, charging, and lighting.Uncertainty estimation is needed because energy demand and supply vary intrinsically.
- Caveats: Urban ML applications face data gaps because high-quality datasets rarely cover smaller cities or the full spectrum of building types.The paper identifies ML-based extraction from satellite imagery as one possible way to obtain broader data.
4 Industry
Industry is a major source of difficult-to-eliminate greenhouse gas emissions, and machine learning offers opportunities across supply chains, production, maintenance, materials, and energy use. These opportunities depend on data, adjustable processes, information sharing, aligned incentives, and attention to rebound effects and adoption barriers.
- Overview: Industrial production, logistics, and building materials are leading causes of difficult-to-eliminate greenhouse gas emissions, while the sector generates extensive operational data.Sensors, QR codes, image recognition, and cloud-based infrastructure increase the availability of industrial data.
- Overview: ML has the greatest potential when processes have accessible high-quality data, adjustable objectives, data-sharing incentives, and incentives aligned with emissions reduction.The paper identifies these conditions as prerequisites for impactful industrial applications.
- Overview: Machine learning can streamline supply chains, improve production quality, predict breakdowns, optimize HVAC, and prioritize clean electricity over fossil fuels.Efficiency gains may increase total production and emissions through the Jevons paradox unless incentives constrain overall emissions.
- Caveats: Industrial decarbonization remains constrained by misaligned incentives, expensive equipment, proprietary data, and uncertain outcomes from efficiency improvements.The paper highlights policy mechanisms such as carbon pricing, subsidies, and penalties as ways to align production incentives with emissions reduction.
- Optimizing supply chains: Demand forecasting can reduce overproduction and excess inventory, whose production, shipment, and warehousing generate substantial industrial emissions.Global excess inventory was estimated at about $8 trillion in 2011, while average corporate sales estimates diverged from actual sales by 40%.
- Optimizing supply chains: ML can reduce food waste through logistics optimization, addressing 1.3 billion metric tons of food lost or wasted globally each year.Waste occurs at different supply-chain stages in developing and industrialized countries.
- Improving materials: ML-enabled generative design, process redesign, and materials discovery could reduce emissions from cement, steel, chemicals, and ammonia production.Approaches include reducing carbon-intensive material use, simulating 3D printing, identifying catalysts, and developing low-emission concrete or ammonia pathways.
- Industrial operations: Predictive maintenance, digital twins, interpretable ML, and demand-response optimization can improve equipment operation and industrial electricity use.These methods can model machinery wear, test processes virtually, and shift electrical loads toward lower-carbon operation.
5 Farms & Forests
Farms and forests are major sources and stores of greenhouse gases, making localized land-management interventions an important application area for machine learning. The paper highlights precision agriculture, remote sensing, forest monitoring, and afforestation while emphasizing risks from efficiency-driven rebound effects and stakeholder complexity.
- Better land management and agriculture could provide about a third of global greenhouse-gas emissions reductions.
- Forests and peatlands: Machine learning can monitor forests and peatlands, predict fire risk, support sustainable forestry, and identify suitable afforestation sites.
- Carbon monitoring: Satellite and hyperspectral data can estimate carbon stocks and emissions, but coarse resolution and spatial-temporal gaps limit precise tracking.
- Precision agriculture: Machine learning can support precision agriculture through irrigation, disease and weed detection, soil sensing, yield prediction, and crop-demand modeling.
- Constraints: Efficiency gains in agriculture and forestry may increase harvesting or demand, so effective applications require policy, regulation, and coordination across stakeholders.
- Reducing deforestation: Only 17% of the world’s forests are legally protected, while deforestation contributes approximately 10% of global greenhouse-gas emissions.
6 Carbon Dioxide Removal
Because eliminating emissions entirely is difficult and atmospheric greenhouse gases persist, carbon dioxide removal may be important for climate goals. The paper surveys potential ML roles in direct-air-capture materials and processes, geological storage characterization, and sequestration monitoring, while noting that many opportunities remain speculative.
- Removal pathways: Plant-, biomass-, mineral-, and ocean-based removal methods can reach large scales but may create land-use or environmental impacts and may not benefit substantially from machine learning.
- Direct air capture: Direct air capture uses sorbents to extract carbon dioxide from exhaust, industrial processes, or ambient air while requiring little land.
- Direct air capture: Machine learning could accelerate sorbent materials discovery and process engineering to improve reusability and uptake while reducing energy requirements.
- Geological sequestration: Machine learning can help identify storage sites, characterize subsurface geology, and monitor active sequestration facilities.
- Carbon dioxide removal may have a critical role because humanity faces limits on safe future emissions and difficulties eliminating emissions entirely.
- Outlook: Many carbon-dioxide-removal applications are speculative, but a growing industry may generate more data and opportunities for machine learning.
7 Climate Prediction
Climate models underpin predictions and decisions, but climate science faces both data-rich opportunities and irreducible data limits. The paper identifies machine learning strategies that can improve accuracy, resolution, computational cost, and ensemble forecasting, while stressing that further model development remains necessary.
- Climate models inform government decisions, climate-risk calculations, and estimates of solar-geoengineering impacts.
- Opportunities: New satellite observations and simulated datasets create petabytes of climate data, while expensive simulations motivate faster machine-learning approaches.
- Opportunities: Machine learning is useful when data are plentiful but difficult to model statistically, or when established physical models are too computationally expensive.
- Data limits: Climate prediction remains data-limited because Earth can generate at most one year of new climate data per year, requiring physical laws in existing models.
- Accelerating climate models: Deep neural networks emulating high-resolution cloud simulations produced similar results at a fraction of the cost and remained stable in a simplified global model.
- Open challenges: Further work is needed to identify additional replaceable model components, optimize them, and automate their training workflows.
- Forecasting: Machine learning can combine or downscale climate-model outputs to create higher-resolution ensemble predictions and localized risk scenarios.
8 Societal Impacts
Climate change creates both gradual stresses and acute disruptions for ecosystems, infrastructure, and societies. The paper organizes machine-learning opportunities around recurring adaptation needs, including risk identification, data annotation, and resource exchange.
- Adaptation needs: Climate impacts include gradual ecological and socioeconomic stresses alongside severe disruptions such as food shortages and disasters.
- Adaptation needs: Recurring machine-learning needs for adaptation include identifying high-risk areas, extracting actionable information from raw data, and sharing resources and information.
- Discussion: The section concludes that effective adaptation requires both recurring machine-learning capabilities and attention to domain-specific needs across local-to-global projects.
- Ecosystems: Ecosystem monitoring can use remote sensing, simulations, sensor networks, and marine robotics to prioritize at-risk environments and support adaptation science.
- Ecosystems: Biodiversity monitoring can infer species counts from camera traps and aerial imagery, while machine learning can assess ecological interventions and help prevent poaching.
- Infrastructure and society: Climate stress threatens roads, buildings, power lines, pipelines, water access, food security, and supply chains, requiring resilience and recovery planning.
- Infrastructure and society: Machine learning can support migration prediction and health-risk assessment as climate change increases displacement and exposure to heat waves and deteriorating air quality.
9 Solar Geoengineering
Solar geoengineering presents machine-learning opportunities in design, emulation, control, and impact assessment, but it also involves substantial side effects, uncertainty, and governance challenges. The paper emphasizes that objectives and harms must be evaluated carefully before considering deployment.
- Overview: Solar geoengineering proposals alter Earth’s heat balance but raise potential side effects and governance challenges.
- Risks and governance: Solar geoengineering cannot simply reverse climate-change effects, may produce termination shock, and can create moral and distributional risks.
- Methods: Marine cloud brightening, cirrus thinning, and stratospheric aerosol injection are identified as primary candidate methods.
- Methods: Stratospheric sulfate injection is a leading candidate because of its economic and technological feasibility and the availability of volcanic-eruption temperature data.
- Design and modeling: Machine learning may help explore aerosol designs, quantify uncertainty, constrain model parameters, and reduce the computational burden of climate-model simulations.
- Impact assessment: Evaluating geoengineering interventions requires objectives that account for harms to people, ecosystems, and society rather than climate variables alone.
- Discussion: The section outlines technical challenges in implementing and evaluating solar geoengineering and calls for further machine-learning contributions.
10 Individual Action
Machine learning can help individuals reduce emissions by identifying high-impact behaviors, tailoring interventions, and automating energy decisions. These tools rely on personal, household, and behavioral data to make climate action more understandable and actionable.
- Behavior and emissions: Machine learning can identify behaviors that meaningfully reduce individual carbon footprints and inform people about which actions matter most.
- Behavior and emissions: Natural language processing can extract flights and purchased goods from personal records to estimate an individual’s carbon footprint.
- Household interventions: Models can predict emissions across transportation, energy, water, waste, food, goods, and services, enabling interventions targeted to high-emissions households.
- Household interventions: Energy disaggregation identifies appliance-level electricity use, including previously unnoticed consumption from devices on standby.
- Energy automation: Real-time marginal-emissions forecasts can help schedule activities such as electric-vehicle charging when emissions and prices are lowest.
- Behavior change: Modeling preferences and characteristics improved enrollment in an energy-savings program by 2-3x.
- Behavior change: Machine learning and natural language processing can personalize policy information and suggest interventions that reduce psychological distance to climate impacts.
11 Collective Decisions
Climate action requires collective decisions among governments, organizations, businesses, and communities under complex incentives, trade-offs, and uncertainty. Machine learning can augment behavioral models, policy analysis, data gathering, optimization, and market design, with interpretability and fairness especially important.
- Collective action: Collective climate decisions span treaties, carbon markets, resilient infrastructure, and community energy projects involving multiple stakeholders.
- Behavioral modeling: Agent-based models simulate people’s and organizations’ actions and interactions, including low-carbon technology adoption.
- Behavioral modeling: Game theory, multi-agent reinforcement learning, and mechanism design can model cooperation and design incentives for mutually beneficial outcomes.
- Policy analysis: Policy analysis can evaluate past policies and future alternatives using statistics, economics, operations research, and machine-learning tools.
- Policy analysis: Machine learning can provide policy-relevant data by identifying emissions sources, traffic patterns, infrastructure risks, and information in financial disclosures.
- Decision tools: Machine learning can support simulations, integrated assessment models, optimization, and multi-criteria decision-making for complex policy trade-offs.
- Markets: Market-based climate strategies require analysis of distributional effects, with clustering and supervised learning offering tools for more equitable designs.
- Discussion: Because climate decisions involve complexity, scale, and fundamental uncertainty, interpretable and fair machine-learning methods may be particularly important.
12 Education
Education supports sustainable development and climate adaptation, while AI and ML can expand access, personalize learning, and help learners connect actions with climate impacts.
- Education improves quality of life, informs decisions, trains innovators, and provides skills for adapting to climate change.
- AI and ML can broaden educational access, personalize teaching, and support instruction when teachers have limited time.
- Intelligent tutoring systems adapt learning activities to individual needs or collaborative-learning contexts using methods including bandits and LSTMs.
- Scalable adaptive online courses and computational teaching guidance can improve educational outcomes for large groups of learners.
- Climate-focused educational activities can connect personal and collective actions to global climate impacts and support climate-friendly choices.
13 Finance
Climate change creates substantial financial risks that are difficult to forecast, motivating climate analytics and broader applications of ML in climate finance. The paper identifies prediction, optimization, risk modeling, and deployment opportunities while emphasizing cross-field collaboration and practical action.
- Climate change threatens global assets measured in the trillions of dollars, but its effects on individual stock prices are difficult to forecast.
- Climate investment directs capital toward low-carbon assets through green indexes and carbon-neutral portfolios.
- Climate analytics predicts climate change’s financial effects by analyzing portfolios, funds, and companies to identify climate-risk exposure.
- ML opportunities include improving portfolio optimization, modeling climate-risk variables, developing climate factors, and identifying climate-risk exposure in company reports.
- ML can support climate mitigation and reduce climate-related financial impacts on society through financial-sector applications.
- The paper presents ML as one component of broader cross-field solutions, alongside monitoring, scientific discovery, optimization, and hybrid physical modeling.