PaperlayerView exact revision on arXiv

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

arXiv:1906.05433v2cs.CYcs.AIcs.LGstat.ML

TL;DR

Climate change demands mitigation and adaptation across diverse systems, while the machine learning community lacks a concerted effort identifying where its tools can help. This paper surveys high-impact applications through cross-field consultation and highlights research and engineering opportunities spanning mitigation, adaptation, and enabling tools.

  • Problem

    Climate change is intensifying, while machine learning practitioners lack guidance on where their tools can best support mitigation and adaptation across diverse domains.

    Source: Introduction · pp. 1–2

  • Method

    The paper provides a cross-disciplinary overview of high-impact machine learning applications for climate mitigation, adaptation, and enabling strategies.

    Source: Introduction · p. 2

  • Results

    The paper identifies numerous climate applications where machine learning may deliver high impact, including climate-model improvement, ensemble prediction, and localized risk forecasting.

    Source: Introduction · p. 2, §7.1.2 · p. 37, §7.1.3 · p. 38

  • Takeaways & Limitations

    The paper frames climate change as a broad opportunity for machine learning research and engineering focused on urgent societal problems.

    Source: Introduction · p. 1, Who is this paper written for? · p. 2

    The paper’s application flags reflect subjective evaluation and its coverage is necessarily incomplete, with some applications potentially overlooked or dismissed erroneously.

    Source: How to read this paper · p. 4

Abstract

from arXiv · show

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

Climate change demands mitigation and adaptation across many sectors, and this paper maps high-impact opportunities for machine learning through engineering, research, and collaboration with domain experts. It targets researchers, practitioners, organizations, and governments while emphasizing that ML is neither definitive nor sufficient without societal action.

Introduction

  • Climate change is intensifying, while mitigation and adaptation require coordinated changes across energy, transport, buildings, industry, land use, resilience, and disaster management.The paper frames this diversity of challenges as an opportunity for machine learning to have impact.Source: Introduction · p. 1
  • The paper identifies high-impact machine-learning applications for climate mitigation, adaptation, and enabling tools, informed by consultation with experts across fields.It addresses ML practitioners seeking direction and fields seeking ML expertise.Source: Introduction · p. 2

Who is this paper written for?

  • The recommendations address researchers, entrepreneurs, corporate leaders, and governments through opportunities spanning conceptual innovation, deployment, efficiency, public services, data, and infrastructure.Examples include interpretable ML for climate models, solar forecasting, energy reduction, supply-chain optimization, precision agriculture, transportation, building assessment, and disaster management.Source: Who is this paper written for? · p. 2

How to read this paper

  • The paper flags applications as High Leverage, Long-term, or Uncertain Impact, but presents these labels as non-definitive judgments and acknowledges omissions and possible errors.High Leverage marks ML-suited bottlenecks; Long-term marks primary impact after 2040; Uncertain Impact marks uncertain emissions effects or potential side effects such as rebound effects.Source: How to read this paper · p. 4

A call for collaboration

  • Effective climate applications require cross-field collaboration, accessible tools and data, and adoption by intended users, while ML alone cannot solve climate change or replace societal action.Domain experts help avoid irrelevant, oversimplified, risky, or unnecessarily complex solutions, and useful code should integrate with platforms users already employ.Source: A call for collaboration · p. 4

1 Electricity Systems

Electricity systems produce about a quarter of human-caused greenhouse gas emissions annually, while electrification will increase demand for low-carbon power. Machine learning can support the transition through clean-energy development, forecasting, optimization, management, and monitoring, in collaboration with experts across relevant fields.

Electricity Systems

  • About a quarter of human-caused greenhouse gas emissions comes from electricity systems each year, while electrification across buildings, transportation, and other sectors will increase demand for low-carbon electricity.Reducing emissions requires rapid expansion of low-carbon sources such as solar, wind, hydro, and nuclear, alongside phasing out carbon-emitting sources.Source: §1 · p. 5
  • Emissions from existing CO2-emitting power plants must also decline because the transition to low-carbon electricity will not happen overnight.These changes must be implemented across countries and contexts because electricity systems are ubiquitous.Source: §1 · p. 5
  • Machine learning can inform electricity-system research, deployment, and operation by accelerating clean-energy development, improving demand and clean-energy forecasts, optimizing and managing systems, and enhancing monitoring.These contributions require varied ML paradigms and close collaboration with electricity-industry experts and fields including operations research, electrical engineering, physics, chemistry, and social sciences.Source: §1 · p. 5

1.1 Enabling low-carbon electricity

Machine learning can reduce emissions from electricity systems by improving forecasting, scheduling, flexible demand, storage, materials discovery, and the integration and operation of low-carbon generation. It can support both variable sources, whose output fluctuates, and controllable sources, while helping develop future technologies such as fusion.

1.1.1 Variable sources

  • ML can reduce emissions from variable electricity by improving forecasting, scheduling, control, storage, and electricity markets that accommodate variable generation and flexible demand.Variable generation must be balanced by controllable resources; ML can reduce reliance on fossil-fuel spinning reserves and support storage-based, lower-carbon systems.Source: §1.1 · p. 5, §1.1.1 · p. 6
  • ML forecasting should combine historical, physical, weather, image, and video data across time and space while quantifying uncertainty and optimizing system goals such as scheduling costs or GHG emissions.Domain-specific physics, climate modeling, interpretability, and automated visualization can make forecasts more useful for operators and longer-term planning.Source: §1.1.1 · p. 7

Improving scheduling and flexible demand

  • Electricity scheduling is slow and complex, but ML can speed optimization, improve solutions, enable real-time balancing, and support decentralized control through storage, flexible demand, and emissions-aware prices.These approaches include approximation, learned initialization, reinforcement learning, local control algorithms, and market mechanisms, but require testing and scaling on real or simulated systems.Source: Improving scheduling and flexible demand · pp. 7–8
  • ML can accelerate materials discovery for solar fuels and batteries by combining heuristics, experimental data, physics, optimization, and reasoning to design and characterize energy materials.Relevant methods include supervised learning, active learning, generative models, and physics-based modeling, although moderately sized datasets and learning physical principles remain challenges.Source: Improving scheduling and flexible demand · p. 8

Additional applications

  • ML can advance variable generation by optimizing solar and wind operation, detecting faults, placing turbines, controlling farm batteries, and integrating rooftop solar into sparsely monitored distribution grids.Applications include reinforcement learning, Bayesian optimization, graphical models, computer vision, matrix completion, and deep neural networks.Source: Additional applications · p. 9

Managing existing technologies

  • For controllable low-carbon sources, ML can guide deployment, maintain geothermal, hydropower, and nuclear fission facilities, and accelerate fusion through experimental design and monitoring of plasma disruptions.Fusion applications must address rapid instabilities, reactor-specific characteristics, and the combination of simulated and experimental data; fusion currently consumes more energy than it produces.Source: Managing existing technologies · pp. 9–10, §1.1.2 · p. 9

1.2 Reducing current-system impacts

Machine learning can reduce current electricity-system impacts by cutting fossil-fuel emissions, preventing avoidable delivery losses, and flexibly managing demand based on grid emissions. These efforts should support, not delay, the transition to low-carbon electricity and be evaluated for their system-wide greenhouse-gas effects.

1.2 Reducing current-system impacts

  • ML can reduce fossil-fuel emissions by detecting methane leaks, optimizing fuel transport and power-plant operations, and managing CO2 storage, but projects require system-impact analysis to avoid delaying decarbonization.Existing work uses sensor and satellite data for proactive pipeline maintenance and leak detection.Source: §1.2 · p. 10

Reducing system waste

  • ML can prevent avoidable electricity-delivery losses by using predictive maintenance to recommend proactive grid upgrades, building on LSTM [136], bipartite-ranking, and neural-network-plus-clustering approaches.Future methods should improve and localize these techniques for different contexts.Source: Reducing system waste · p. 10

Modeling emissions

  • Flexible demand and storage management require accurate marginal emissions factors, which current systems estimate or forecast in real time or across multiple days using regression and ensemble models.Future work should improve these forecasts, model curtailment, and quantify uncertainty so load shifting decreases rather than increases emissions.Source: Modeling emissions · pp. 10–11

1.3 Ensuring global impact

Machine learning can broaden climate impact by improving clean electricity access and translating insights from high-data electricity systems to low-data contexts. These applications can support mitigation and social and economic development across diverse settings [26, 142, 143].

1.3 Ensuring global impact

  • Because electricity research often focuses on near-universal-access, data-rich settings such as the United States, global climate impact requires considering places lacking these attributes.Source: §1.3 · p. 11

Improving clean energy access

  • Transfer learning and data-efficient methods can translate grid insights from high-data to low-data settings, including weakly supervised models constrained by physical or domain-specific knowledge.This is enabled in part because electric grids share underlying system physics, despite limited data collection and sharing in many countries.Source: Improving clean energy access · p. 11
  • ML can generate electricity-system information where data are scarce by mapping grids, estimating power-plant CO2 emissions, and modeling consumption from cellular-network data.These approaches use satellite imagery, graph search, and regression to address settings with few grid sensors.Source: Improving clean energy access · p. 11

1.4 Discussion

Electricity-system ML offers climate opportunities but also domain-specific pitfalls, requiring close collaboration, interpretable and physics-aware models, and methods suited to data-scarce grids.

1.4 Discussion

  • Reducing oil-and-gas emissions through innovation could instead increase emissions by making them cheaper to emit.Electricity-system ML therefore requires close collaboration with decision-makers and practitioners in electrical engineering, natural sciences, and social sciences.Source: §1.4 · p. 12
  • Interpretable and hybrid ML models can help stakeholders apply models while incorporating system physics, domain-specific goals, and existing domain knowledge [151] [152] [58] [59] [60].These approaches are presented as ways to address domain-specific nuances in real-world electricity systems.Source: §1.4 · p. 12
  • Applying data-driven insights to modern electric grids that are not data-abundant may be the next grand challenge for ML in electricity systems.Such grids can be data-driven without providing abundant data for model development.Source: §1.4 · p. 12

2 Transportation

Transportation produces about a quarter of energy-related CO2 emissions and remains difficult to decarbonize, but machine learning can support reductions through improved data, planning, vehicle efficiency, electrification, and modal integration. Its effects depend on infrastructure, behavior, and strong public policies, especially for shared mobility and autonomous vehicles.

2 Transportation

  • ML can decarbonize transportation through better data, planning, engineering, infrastructure, operations, and policy-relevant information, provided interventions align with strong public policies.Transportation remains hard to decarbonize because of fuel energy-density requirements and controversial effects on end-users.Source: §2 · p. 13, §2.5 · p. 19

2.1 Reducing transport activity

  • ML can reduce transport activity by forecasting demand, estimating mobility patterns, optimizing freight consolidation and routing, and reducing congestion-related fuel use in airports.Applications include origin-destination estimation, transit ridership forecasting, runway-demand and taxi-time prediction, and freight-disruption planning.Source: §2.1 · p. 13, Understanding transportation data · pp. 14–15

Understanding transportation data

  • ML converts sparse transportation data into traffic counts, road classifications, satellite-based vehicle estimates, missing-data imputations, emissions simulations, and demand forecasts for infrastructure planning.These methods draw on computer vision, sensors, satellite imagery, and novel mobility data such as smart cards and mobile phones.Source: Understanding transportation data · pp. 13–14

2.5 Discussion

  • Shared mobility, autonomous vehicles, telepresence, and additive manufacturing have uncertain emissions effects, so ML and policy must prevent higher travel demand, vehicle-miles, or carbon-intensive technological lock-in.Potential benefits include pooled rides, public-transit integration, electric shared vehicles, reduced freight, and substituted passenger travel, but outcomes depend on incentives and energy sources.Source: Understanding transportation data · pp. 15–16, §2.5 · p. 19

2.2 Improving vehicle efficiency

  • ML can improve vehicle efficiency through engine and hybrid power-management design, aerodynamic surrogate modeling, lighter components, and manufacturing-process failure detection and material design.Autonomous vehicles could reduce congestion and energy use, but increased road traffic may nullify those gains; freight platooning and delivery robots offer more targeted savings.Source: Designing for efficiency · pp. 16–17

2.3 Alternative fuels and electrification

  • EVs are a primary decarbonization pathway, and ML supports charging, congestion management, vehicle-to-grid control, battery management, fault detection, and battery-life prediction.Vehicle-to-grid systems may reduce electricity-generation emissions but could face consumer reluctance, while battery models need real-world operational and environmental data.Source: §2.3 · p. 17
  • Alternative fuels such as electrofuels, biofuels, and hydrogen are important for aviation, long-distance road transport, and shipping, and ML may improve their research and development.Their adoption is constrained by cost, land use, infrastructure compatibility, and the early stage of development.Source: §2.3 · pp. 17–18

2.4 Modal shift

  • Modal shift toward public transit, rail, and water can lower emissions, while ML improves mode-choice analysis, multimodal coordination, transit reliability, freight interoperability, and low-carbon service operations.Applications include transit and freight arrival-time prediction, bike-sharing rebalancing, integrated travel-time estimates, and enforcement against truck overloading.Source: §2.4 · p. 18, Passenger preferences · pp. 18–19

3 Buildings & Cities

Buildings and cities offer substantial opportunities to reduce greenhouse-gas emissions, but heterogeneous, long-lived infrastructure and limited data complicate action. Machine learning can support building management, urban planning, and coordinated transitions to low-carbon cities while requiring attention to rebound effects, privacy, equity, and geographic bias.

3 Buildings & Cities

  • Buildings account for a quarter of global energy-related emissions, while combined efficiency measures could reduce existing-building emissions by up to 90% and often save costs.Buildings can also provide services such as air conditioning as climate change increases heat-wave risks.Source: §3 · p. 20

Modeling building energy

  • Building heterogeneity and long lifespans require context-specific strategies combining new efficient construction, retrofits, urban planning, incentives, and standards.Buildings differ by age, construction, usage, ownership, and energy context, so a strategy effective in one building may not suit another.Source: §3 · p. 20

3 Buildings & Cities

  • Machine learning supports building managers and policy makers by tailoring building strategies, optimizing energy use, informing urban planning, and guiding cities toward low-carbon futures.These applications span individual building control, city-scale data analysis, and coordinated urban transitions.Source: §3 · p. 20

3.1 Optimizing buildings

  • Machine learning can model building energy consumption and optimize smart-building operations, including forecasting demand, detecting appliances and faults, and controlling HVAC systems.Deep reinforcement learning achieved a scalable 20% energy reduction for HVAC control using only three sensors.Source: §3.1 · p. 20, Modeling building energy · pp. 20–22

3.2 Urban planning

  • Smart-building systems can adapt consumption to occupancy, electricity prices, grid signals, and learned preferences, but rebound effects can add 10–20% energy use and create privacy, security, and embodied-energy concerns.Public policies are needed to complement individual building managers and ensure that cost optimization produces emissions reductions.Source: Modeling building energy · pp. 21–22

Modeling energy use across buildings

  • Urban Building Energy Models and other machine-learning methods can predict energy use, identify retrofit potential, and scale forecasts across districts despite sparse energy-consumption data.A hybrid neural-network and physical model produced precise campus-wide forecasts accounting for inter-building dynamics and urban microclimate factors.Source: §3.2 · p. 22, Modeling energy use across buildings · p. 23

3.3 The future of cities

  • Smart-city machine learning can integrate hard- and soft-sensing data, preprocess real-time sensor streams, and combine heterogeneous sources for data-driven climate policy, while privacy remains paramount.Relevant inputs include traffic sensors, smartphones, geotagged images, credit-card payments, and social-media activity.Source: §3.3 · p. 24

Low-emissions infrastructure

  • Coordinated low-emissions infrastructure can link heating, cooling, solar generation, electric-vehicle charging, and public lighting, but current smart-city research is concentrated in wealthier regions and the Global North.Strategies can be transferred by clustering cities according to climate-relevant dimensions, while infrastructure models may support low-carbon neighborhoods without advanced smart-city technologies.Source: Low-emissions infrastructure · pp. 24–25

4 Industry

Machine learning could reduce industrial greenhouse-gas emissions across supply chains, materials, production, maintenance, and energy use, provided that data, adjustment opportunities, information sharing, and incentives align. Its impact is constrained by proprietary or poor-quality data, costly errors, misaligned incentives, and the need for robust, safe, interpretable, and reproducible systems.

4 Industry

  • ML can reduce industrial GHG emissions by streamlining supply chains, improving production quality, predicting breakdowns, optimizing heating and cooling, and prioritizing clean electricity, but efficiency gains may increase production without emissions incentives.Industry is especially promising because it generates extensive data through sensors and other mechanisms, alongside affordable cloud storage and computing.Source: §4 · p. 26, Figure 4 · p. 26

4.1 Optimizing supply chains

  • ML can optimize supply-and-demand forecasts, identify lower-carbon products, and select lower-emission shipping routes, but climate benefits require firms’ financial incentives to align with mitigation.Better forecasting could reduce overproduction and excess inventory, although just-in-time and online fulfillment can also increase smaller, less-efficient road shipments.Source: §4.1 · p. 27

Recommender systems

  • ML could direct consumers and firms toward climate-friendly products and reduce food waste through demand forecasting, route optimization, refrigeration improvements, and spoilage-detection sensors.Recommender systems require life-cycle emissions data spanning materials, production, shipping, use, and disposal, plus company data sharing.Source: Recommender systems · p. 27

4.2 Improving materials

  • ML combined with generative design, 3D-printing simulation, and materials databases could reduce carbon-intensive material use and help discover lower-emission concrete, catalysts, and ammonia-production materials.Cement and steel production together account for over 10% of global GHG emissions, while ammonia production uses around 2% of global energy consumption.Source: §4.2 · p. 28

4.3 Production and energy

  • ML can reduce electricity consumption and optimize HVAC and industrial processes for low-carbon energy, but adoption depends on incentives, affordable retrofits, process-specific data, and accessible tools for factory workers.DeepMind used reinforcement learning to optimize cooling-center power usage effectiveness for Google servers, lowering HFC emissions and cooling costs.Source: §4.3 · p. 28

Predictive maintenance

  • Predictive maintenance and digital twins can prevent equipment failures, production waste, leaks, and unsafe scenarios while enabling virtual testing and more GHG-efficient industrial processes.ML can model machinery wear, identify broken or failing machines before damaged products are produced, and predict pipes liable to leak HFCs or natural gas.Source: Predictive maintenance · p. 29

4.4 Discussion

  • Industrial ML research must address misaligned incentives, proprietary or low-quality data, costly mistakes, and managerial risk aversion by delivering robust, safe, interpretable, and reproducible systems.Practitioners estimate that 60-70% of industrial data goes unused.Source: §4.4 · p. 29

5 Farms & Forests

Human land use causes about a quarter of global greenhouse-gas emissions, but better land management and agriculture could provide about a third of emissions reductions. Machine learning can support this effort through remote sensing, precision agriculture, peatland and forest monitoring, fire management, and deforestation reduction.

5 Farms & Forests

  • Land use causes about a quarter of global GHG emissions, while machine learning could support better management of agriculture, forests, peatlands, and fires.Deforestation, unsustainable agriculture, cattle, and rice farming release carbon and methane, while climate change increases emissions from permafrost, peatlands, and fires.Source: §5 · p. 30

5.1 Remote sensing of emissions

  • Real-time GHG maps could quantify agricultural and forestry emissions, guide regulations and incentives, and improve enforcement by pinpointing emission sources.Hyperspectral satellites can estimate several gases, but coarse resolution and data gaps leave ML-based fusion with high potential impact and no clear results yet [405, 406, 407, 408].Source: §5.1 · pp. 30–31

5.2 Precision agriculture

  • Precision agriculture uses ML, robotics, sensing, and prediction to adapt farming to local conditions while improving irrigation, crop management, and potentially emissions.Applications include mechanical weeding, targeted pesticide use, disease and weed detection, soil sensing, yield and demand prediction; efficiency gains may not reduce emissions because of the Jevons paradox, and broader paradigm shifts may be needed.Source: §5.2 · pp. 31–32

5.3 Monitoring peatlands

  • Peatlands cover 3% of land yet hold twice the carbon in all forests, making ML-based monitoring, carbon-stock estimation, and fire-risk prediction especially valuable.Protecting peatlands from drainage and drought is essential, and a 1997 Indonesian peat fire released emissions comparable to 20–50% of global fossil-fuel emissions that year.Source: §5.3 · p. 32

5.4 Managing forests

  • ML can scale forest carbon-stock estimation from LiDAR and satellite imagery and support afforestation by locating planting sites, monitoring health, and assessing weeds.Afforestation must avoid displacing farmland or peat bogs and should respect local conditions and native species to prevent higher emissions or biodiversity loss.Source: §5.4 · p. 32

Managing forest fires

  • ML can forecast fire risk, model fire progression, detect deforestation, and support selective harvesting, but prevention and efficiency can backfire through ecological damage or the Jevons paradox.Tools include drought and canopy-water forecasting, reinforcement learning for fire progression, remote sensing and acoustic detection of illegal cutting, and harvest-planning systems.Source: Managing forest fires · p. 33

5.5 Discussion

  • Meaningful ML applications face a globalized problem, localized actions, diverse stakeholders, and fragmented access to relevant data.Agriculture spans large farms, small farmers, equipment manufacturers, and chemical companies with differing interests and data, making stakeholder coordination a practical challenge.Source: §5.5 · p. 33

6 Carbon Dioxide Removal

CO2 removal may be critical because emissions cannot be eliminated entirely, but the industry remains immature and many ML applications are speculative. Promising ML opportunities include improving direct-air-capture components, characterizing storage resources, and monitoring sequestration sites.

6 Carbon Dioxide Removal

  • Meeting climate goals may require net-negative emissions, but CO2 removal remains immature and candidate biomass, mineral, and ocean methods may have land-use or environmental drawbacks.These natural-process-based methods are promising at scale but are unlikely to benefit significantly from ML.Source: §6 · p. 34

6.1 Direct air capture

  • Direct air capture uses sorbents and heat-powered processes to extract CO2 from exhaust or ambient air, while ML could improve uptake, reusability, efficiency, and component durability.Potential applications include materials discovery, process engineering, corrosion-resistant components, and geometry optimization for air-sorbent contact.Source: §6.1 · p. 34

6.2 Sequestering CO2

  • ML could identify and characterize CO2 storage sites, infer subsurface flow and capacity, quantify uncertainty, and monitor leaks at active sequestration facilities.Potential inputs and methods include seismic traces, noisy sensors, convolutional image-to-image regression, and computer vision.Source: §6.2 · p. 35

6.3 Discussion

  • ML opportunities in CO2 removal include developing component materials, characterizing geologic resources, and monitoring underground CO2, although many applications remain speculative.Industry growth should create more data and opportunities for ML approaches.Source: §6.3 · p. 35

7 Climate Prediction

Climate prediction relies on physics-based climate models, while expanding observations, simulations, and faster ML create opportunities to improve accuracy, reduce computational cost, and make forecasts more actionable. The section emphasizes integrating ML with existing scientific models and domain expertise, especially for data-limited systems and extreme events.

7 Climate Prediction

  • ML can improve climate prediction when data are abundant or existing physical models are too computationally expensive, while climate models remain essential for representing Earth’s physics.Climate models support government decisions, climate-risk assessment, and solar-geoengineering analysis.Source: §7 · p. 36, §7.1 · p. 37

7.1.1 Data for climate models

  • Curated geoscience datasets, including EnviroNet, IS-GEO, and ExtremeWeather, can expand data-driven climate modeling through collaboration between modellers and domain experts.ML may also optimize data collection, while practitioners should address fitting and over-fitting in climate data.Source: §7.1.1 · p. 37

7.1.2 Accelerating climate models

  • Neural networks can emulate costly cloud simulations with similar results at a fraction of the cost, potentially improving global climate-model accuracy and efficiency.ML is also promising for uncertain ice-sheet and sea-level processes, including snow and sea-ice reflectivity, ocean heat mixing, and grounding-line migration.Source: §7.1.2 · pp. 37–38

7.1.3 Working with climate models

  • ML can exploit relationships among climate variables and combine predictions from 20+ climate models, including through online learning and random forests for high-resolution ensemble forecasts.A proposed Climate Modeling Alliance model would continuously learn from data and high-resolution simulations and use Julia to improve accessibility to developers and ML libraries.Source: §7.1.3 · p. 38

7.2 Forecasting extreme events

  • ML may bridge difficult weeks-to-months prediction gaps, while weather applications already use methods such as random-forest ensembles to improve hail forecasts.Weather models have far more test data and earlier ML adoption than climate models.Source: §7.2 · pp. 38–39

7.2.1 Storm tracking

  • Deep learning has classified, detected, and segmented rare extreme events such as cyclones, atmospheric rivers, and tornadoes in historical climate datasets.Future tools should support more event types, online climate-model use, labeled future-event datasets, and uncertainty quantification.Source: §7.2.1 · p. 39

7.2.2 Local forecasts

  • ML downscales coarse 10–100 km forecasts into local risk scenarios, including localized flooding and temperature scores that can inform insurance, housing, infrastructure, and disaster decisions.Neural networks are effective for predicting local flooding during extreme weather, and Jupiter Intelligence is translating forecasts into localized risk scores.Source: §7.2.2 · p. 39

7.3 Discussion

  • Successful climate ML should remain integrated with scientific models, leverage physical knowledge under limited data, and be developed by close-knit climate and computational-science teams.Promising directions include physics-informed neural networks, neural ordinary differential equations, adaptive meshing, uncertainty quantification, and performance optimization.Source: §7.3 · pp. 39–40

8 Societal Impacts

Machine learning can help societies adapt to climate change by reducing vulnerability, supporting rapid recovery, and strengthening ecosystems, infrastructure, social systems, livelihoods, health, and crisis response. Recurring opportunities include sounding alarms, extracting actionable information, and promoting resource and information exchange, while deployment must address heterogeneity, uncertainty, transparency, and risks to vulnerable populations.

8 Societal Impacts

  • ML supports societal adaptation by sounding alarms, providing actionable annotations, and promoting exchange across ecosystem, infrastructure, and social strategies.These applications range from local infrastructure upgrades and crisis preparedness to global ecosystem monitoring and disease surveillance.Source: §8 · pp. 41–42

8.5 Discussion

  • The section calls on ML researchers, engineers, entrepreneurs, and civil servants to pursue domain-specific methods that deliver tangible societal resilience amid climate change.The authors emphasize that progress toward adaptation is increasingly urgent as climate change affects ecosystems and people worldwide.Source: §8 · p. 42, §8.5 · p. 46

8.1 Ecology

  • ML can monitor ecosystems and biodiversity through remote sensing, simulations, sensor networks, marine robots, camera traps, and citizen-science data, while supporting ecological interventions.Effective systems must personalize predictions across ecosystems and address transfer, heterogeneity, failed sensors, missing data, sampling bias, and dataset shift [516–549].Source: §8.1 · pp. 42–43

8.2 Infrastructure

  • ML can improve infrastructure resilience by identifying vulnerable assets, optimizing long-term investments under uncertain climate futures, and prioritizing predictive maintenance or anomaly detection.Applications include flood-aware road planning, localized warnings, resilient utility components, phased water-infrastructure upgrades, and rapid failure discovery [571].Source: §8.2 · pp. 43–44

8.3 Social systems

  • ML can strengthen social systems by forecasting food insecurity, crop yields, and disease risks, while making food supply chains more robust to unexpected disruptions.Models can link mobile-phone, transaction, and social-media proxies with representative consumption data, and combine imagery, meteorology, historical yields, and biological simulations [578].Source: §8.3 · p. 44

Resilient livelihoods

  • ML can promote resilient livelihoods through resource sharing, mobile money, skill matching, training, and migration forecasting, but systems affecting refugees must prevent surveillance-based harm.Research should develop governance and methodology that let vulnerable populations benefit from data without increasing their risk [605].Source: Resilient livelihoods · pp. 44–45

Assessing health risks

  • ML can measure and mitigate climate-related health risks by sensing heat, water quality, and air pollution and identifying populations most vulnerable to resulting hazards.Transparency and causal estimation are needed when interventions rely on observational or unrepresentative data, to avoid feedback loops that systematically exclude subgroups [607–622].Source: Assessing health risks · p. 45

8.4 Crisis

  • ML can support epidemic management and disaster response through disease surveillance, outbreak forecasting, point-of-care diagnostics, aerial mapping, damage assessment, and social-media information retrieval.These tools can guide data collection, route planning, evacuation, retrofitting, relief delivery, and diagnosis where specialized laboratory equipment is inaccessible [631] [635].Source: §8.4 · p. 45, Managing epidemics · pp. 45–46

9 Solar Geoengineering

Solar geoengineering could moderate some catastrophic climate hazards, but its side effects, governance challenges, irreversibility limits, and termination-shock risk prevent this paper from recommending any technique. The section identifies machine-learning opportunities spanning aerosol design, uncertainty-aware modeling, planetary control, impact assessment, and ethical research support.

9 Solar Geoengineering

  • Solar geoengineering deliberately alters Earth’s heat balance, usually by increasing reflectivity, but cannot fully reverse location-specific climate changes and risks termination shock.The paper does not evaluate or recommend a particular technique, despite established potential to moderate catastrophic climate hazards.Source: §9 · p. 47

Overview

  • Primary approaches include marine cloud brightening, cirrus thinning, and stratospheric aerosol injection, alongside less effective or harder-to-implement alternatives.Stratospheric sulfate injection is attractive because of economic and technological feasibility and volcanic-eruption temperature data, but raises ozone-loss and agricultural-diffusion concerns.Source: Overview · p. 47

9.1 Understanding and improving aerosols

  • Machine learning could accelerate aerosol discovery by optimizing chemical and material properties toward reflective, nonreactive, inexpensive particles that remain aloft.The aerosol design space includes sulfates, zirconium dioxide, titanium dioxide, calcite, and synthetic diamond, but remains far from fully explored.Source: §9.1 · p. 47

Modeling

  • Machine learning could quantify aerosol-model uncertainty and reduce computational costs; Gaussian-process emulation required 350 instead of >100,000 climate-model runs for nine parameter settings.The paper argues that future uncertainty-aware aerosol simulations should cost a fraction of one climate-model run rather than 350.Source: Modeling · pp. 47–48

9.2 Engineering a planetary control system

  • Bayesian, neural-network, and reinforcement-learning methods could enable fast, uncertainty-aware control of aerosol interventions in a nonlinear planetary system [664].Potential applications include deciding where and when to release aerosols, with an initial reinforcement-learning analysis using a neural-network climate-model emulator [664].Source: §9.2 · p. 48

9.3 Modeling impacts

  • Because geoengineering cannot preserve every climate characteristic everywhere, machine learning should estimate harms to people, ecosystems, and society alongside climate-model outputs.Deep neural networks have been used to estimate aerosol effects on human health and agriculture, while regression methods estimate temperature effects on economic production.Source: §9.3 · p. 48

9.4 Discussion

  • Solar geoengineering research faces moral-hazard, distributive, and termination-shock concerns, yet may moderate climate uncertainty, motivating further technical research.The section aims to help the machine-learning community address technical challenges in implementing and evaluating geoengineering.Source: §9.4 · pp. 48–49

10 Individual Action

Machine learning can help individuals reduce climate impacts by identifying high-emission behaviors, personalizing interventions, scheduling energy use, and making behavior change more understandable. These tools can empower consumers to act on their carbon footprints and, if widely adopted, contribute meaningfully to mitigation [404, 675].

10.1 Understanding personal carbon footprint

  • ML can estimate personal and household emissions from behavioral data, then use counterfactuals and tailored interventions to guide carbon-reducing choices.Applications include extracting flights from email, identifying groceries from bills, predicting emissions across household activities, and targeting high-emissions households; many mitigation strategies also provide consumer cost savings.Source: §10.1 · p. 50
  • ML can disaggregate household electricity use and predict real-time marginal emissions, enabling interpretable appliance scheduling and coordinated demand management.Standby power accounts for roughly 8% of residential electricity demand, and reinforcement or multi-agent learning can optimize appliance schedules and lower peak demand.Source: §10.1 · p. 50

10.2 Facilitating behavior change

  • ML can model consumer preferences and demographics to target sustainable interventions, increasing energy-program enrollment by 2-3x.Household carbon footprints vary substantially across geographic regions and demographics, while willingness to pay for avoiding environmental harms differs sharply between groups.Source: §10.2 · p. 51
  • ML can facilitate behavior change through personalized policy insights, psychologically informed nudges, climate visualizations, and gamified energy exploration.Behavior-change programs may cost as little as 3 cents per kilowatt hour saved, compared with 5-6 cents for coal or wind generation and 10 cents for solar.Source: §10.2 · p. 51

10.3 Discussion

  • ML can empower consumers to identify high-emission behaviors, schedule energy use automatically, and understand how to facilitate behavior change.These capabilities support individual climate action despite the perception that personal contributions are dwarfed by other factors.Source: §10.3 · p. 51

11 Collective Decisions

Collective climate decisions involve complex, long-term trade-offs among stakeholders, and machine learning can supplement empirical and mathematical approaches across social modeling, policy analysis, optimization, and market design. These tools can improve evidence and decision support, but interpretable and fair ML remains important because qualitative judgments about norms, values, and equity may not be captured quantitatively.

11.1 Modeling social interactions

  • Agent-based models simulate climate-relevant agent interactions, and ML can learn behavioral rules from data, though data-driven approaches may reduce interpretability.Applications include low-carbon technology adoption, social norms, economic and climate evolution, and migration; game theory, multi-agent reinforcement learning, and mechanism design can study and promote cooperation.Source: §11.1 · p. 52

11.2 Informing policy

  • ML can provide policy data and improve policy evaluation by analyzing emissions, infrastructure, public discourse, prior policies, and causal effects at scale.Policy analysis uses cost-benefit, uncertainty, and multi-criteria decision tools, while causal inference has estimated effects such as temperature on mortality.Source: §11.2 · p. 53, Assessing policy options · p. 54

Assessing policy options

  • Integrated assessment models explore climate-consistent societal pathways, while ML can improve their spatial resolution, uncertainty analysis, computational submodels, and validity assessment.ML can aggregate high-resolution outputs, downscale coarse data, and help solve expensive submodels; optimization methods can also search for Pareto-optimal policy solutions under constraints.Source: Assessing policy options · pp. 53–54

11.3 Designing markets

  • Market-based climate strategies aim to internalize the societal cost of greenhouse-gas emissions, and ML can assess their effectiveness, implementation, and fairness.Carbon taxes and quotas affect existing-market supply and demand, whereas cap-and-trade creates emissions-permit markets whose prices depend strongly on permit supply.Source: §11.3 · p. 54, Predicting carbon prices · p. 54, Assessing market effects · p. 55
  • ML supports carbon-market analysis by predicting prices, identifying price drivers, modeling dynamic pricing and auctions [776–778], and assessing economic and distributional effects.Applications include permit allocation, predicting winners and losers from electricity tariffs, valuing environmental goods, and evaluating the viability of forest-conservation offsets.Source: Predicting carbon prices · p. 54, Assessing market effects · p. 55

11.4 Discussion

  • ML can supplement agent-based models, integrated assessment models, multi-objective optimization, and market design for complex climate decisions, while interpretable and fair methods support equitable use.Climate decisions may ultimately depend on qualitative discussions around norms, values, or equity that quantitative models do not capture.Source: §11.4 · p. 55

12 Education

Education supports sustainable development, climate understanding, adaptation, and sustainable behavior, while AI and machine learning can expand access, personalize learning, and improve teaching. Applications range from intelligent tutoring and scalable educational tools to personalized climate-change learning experiences.

12 Education

  • Education improves climate resilience, informed decision-making, innovation, and sustainable lifestyles, while climate change can reduce educational outcomes through effects on agriculture and household income.Source: §12 · p. 56
  • AI and machine learning can broaden educational access, personalize teaching, and support teachers, extending AIED’s established theory-based approaches with data-driven insights.Source: §12 · p. 56
  • Intelligent tutoring systems adapt instruction using bandits, LSTMs, and reinforcement learning, while NLP, learner analytics, and adaptive learning aim to narrow the gap with human tutors.Source: §12 · p. 56
  • Scalable adaptive courses and computational teaching guidance can improve learning for large groups, but global deployment must address technology, electricity, internet, language, culture, and computing skills.Source: §12 · p. 56

Learning about climate

  • Climate-focused education can connect personal actions to collective impacts and encourage lower energy use, while ML can personalize learning through location-specific extreme-weather visualizations and digital replicas.Source: Learning about climate · pp. 56–57

13 Finance

Climate change creates trillions of dollars in financial risks that current short-term financial systems poorly incentivize investors to forecast. The paper identifies climate analytics and machine learning as tools for assessing exposure, improving investment decisions, and supporting broader climate action through interdisciplinary collaboration and deployment.

13 Finance

  • Climate change threatens global assets measured in the trillions of dollars, while quarterly and yearly financial performance overlooks many medium- and long-term climate exposures.These exposures include physical damage to assets and distribution chains, legislative effects on profits, and indirect supply-and-demand consequences.Source: §13 · p. 58

Climate investment

  • Machine learning can support climate investment through deep-learning portfolio selection and investment timing, although current climate investment strategies primarily target low-carbon assets through green indexes and carbon-neutral portfolios.These strategies are shifting utilities and energy markets toward renewable alternatives viewed as having greater growth potential than oil and gas.Source: Climate investment · p. 58
  • Climate analytics can use natural language processing, econometrics, and machine learning to identify climate risks, hedge exposure, exploit carbon-risk factors, and forecast carbon prices.Potential improvements include climate-risk modeling, climate factors for projecting stock-price variation, portfolio optimization, and identifying direct and indirect exposure in company reports.Source: Climate investment · p. 58

Conclusion

  • Effective climate ML requires domain collaboration and attention to proprietary, sensitive, heterogeneous, and geographically unrepresentative data, with transfer learning, domain adaptation, simulation, and broader data release helping address low-data settings.The paper highlights interpretability, causality, uncertainty quantification, and physics-constrained ML as methodological opportunities, while emphasizing that ML is only one part of interdisciplinary solutions.Source: Conclusion · p. 59
  • The paper’s roadmap asks the machine learning community to learn, collaborate, listen to stakeholders, and deploy solutions where they can have impact in the global effort against climate change.It stresses that both cutting-edge technologies and well-constructed solutions to ordinary problems can matter.Source: Conclusion · p. 59

Related papers

Loading 1906.05433v2…