Intro
L1
·
Quiz
·
Lab
L2
·
Quiz
·
Lab
L3
·
Quiz
·
Lab
L4
·
Quiz
·
Lab
Module Test
AI & Climate · Introduction

AI consumes energy, models the climate, and could accelerate solutions. Its relationship to the planet is ambivalent and urgent.

The same technology that's straining the grid may also be the best tool we have for understanding what to do about it.

A single frontier training run can consume as much electricity as a small town uses in a year. AI-specific data-center demand is already reshaping energy markets in Virginia, Ireland, and Oregon. The environmental cost of AI is no longer theoretical — it's showing up in utility bills and siting debates.

At the same time, AI is one of the most powerful tools humanity has ever built for understanding the climate. Climate models that ran for months now run in hours. Protein folding for next-generation materials runs in seconds. Satellite imagery analyzing deforestation, coral bleaching, and methane leaks at unprecedented scale runs continuously. The technology is both a cost and a hope, often simultaneously.

This course holds both truths. It covers the actual energy footprint of AI (training, inference, cooling, water), the climate and environmental-science applications AI is enabling, the policy questions (carbon accounting, siting, grid impact), and the open question that sits behind all of it: whether AI's climate contribution, in the end, is net-positive or net-negative — and what we can do to tilt it.

If you finish every module, here's who you become:

  • You'll understand the actual energy and water costs of AI — from a single training run to global inference at scale.
  • You'll be able to explain why the same technology straining the grid is also accelerating climate modeling, materials science, and ecosystem monitoring.
  • You'll know how AI is reshaping energy systems through smart grids, demand forecasting, and renewable optimization — and where those gains are real versus overstated.
  • You'll recognize the distributional questions behind AI infrastructure: who bears the emissions, the water draw, and the land use when a data center gets built.
  • You'll become someone who can read carbon accounting debates and siting disputes with the technical grounding to form a defensible position.
  • You'll be able to assess whether a given AI climate application is net-positive or net-negative — and name what would need to change for the answer to shift.
  • You'll leave thinking like a practitioner who holds both truths: AI's environmental cost and its planetary-scale problem-solving potential, without collapsing either one.
Lesson 1 · AI & Climate Science

Climate Models and Machine Learning

How neural networks are rewriting what we can predict — and how fast.
Can a machine trained on atmospheric data do in seconds what supercomputers take days to simulate?

When Google DeepMind's GraphCast system produced a 10-day global weather forecast in under 60 seconds on a single TPU chip — a task that previously required a full supercomputer cluster running for hours — it was not a publicity stunt. It was a demonstration that machine learning had crossed a threshold in atmospheric science.

Published in Science in November 2023, GraphCast outperformed the European Centre for Medium-Range Weather Forecasts (ECMWF) operational model on 90% of tested variables at lead times beyond three days. The implications for climate science were immediate and debated.

Why Traditional Climate Models Hit Limits

General Circulation Models (GCMs) — the physics-based engines behind climate projections from institutions like NCAR, GFDL, and the UK Met Office — divide Earth's atmosphere, ocean, and land surface into three-dimensional grids. Each grid cell exchanges heat, moisture, and momentum with its neighbors according to equations derived from fluid dynamics and thermodynamics.

The computational cost scales brutally with resolution. Halving the horizontal grid spacing requires roughly eight times the compute (a factor of four in horizontal area, times two for the shorter timestep needed for numerical stability). The result: the highest-resolution global climate models in the CMIP6 ensemble run at ~25 km resolution. Many cloud processes operate at scales of meters to kilometers — far below what can be explicitly resolved.

This gap is bridged by parameterization: statistical approximations that represent sub-grid processes. Parameterization schemes are a major source of uncertainty across models. Clouds alone account for roughly half the spread in equilibrium climate sensitivity estimates.

Key Limitation

CMIP6 models have equilibrium climate sensitivity (ECS) estimates ranging from 1.8°C to 5.6°C per CO₂ doubling — an uncertainty range that has barely narrowed since the 1979 Charney Report. Much of that range traces to cloud parameterization disagreements.

What Machine Learning Adds

ML approaches to climate science fall into several distinct categories, each addressing a different bottleneck:

Emulation Training a neural network to replicate the input-output behavior of an expensive physical model component at a fraction of the cost. Columbia's M2LInES project trained ML emulators for ocean mesoscale eddies that run 100× faster than the parameterizations they replace.
Hybrid Modeling Embedding learned components directly inside physics-based models. The ClimSim dataset (released 2023) provides training data specifically for learning sub-grid physics inside GCMs, enabling learned parameterizations to replace handcrafted ones.
Data-Driven Forecasting End-to-end models trained on reanalysis data (ERA5) that learn atmospheric dynamics from observations rather than physics equations. GraphCast, Pangu-Weather (Huawei, 2023), and NVIDIA's FourCastNet all take this approach.
Downscaling Using ML to add local detail to coarse model output — analogous to super-resolution in image processing. NVIDIA's CorrDiff model demonstrated statistical downscaling from 25 km to 3 km resolution for precipitation fields.

GraphCast and Pangu-Weather: A Benchmark Moment

GraphCast uses a graph neural network (GNN) architecture. The atmosphere is represented as a graph where nodes correspond to spatial locations and edges capture relationships. After training on 40 years of ERA5 reanalysis data (1979–2018), the model autoregressively rolls forward in 6-hour steps.

Pangu-Weather (Huawei Cloud) uses a 3D Earth Attention Transformer. Both models achieved headline performance: outperforming ECMWF's Integrated Forecasting System (IFS) at medium-range lead times on RMSE metrics for geopotential, temperature, and wind speed.

The crucial caveat: these ML models are trained on ERA5, which was itself produced by ECMWF's data assimilation system. They are in some sense learning to emulate what physics-based systems already know. For novel climate states — conditions outside the training distribution — their reliability is genuinely unknown.

Real Event — October 2023

Cyclone Michaung made landfall near Chennai, India on December 4, 2023. Researchers at ECMWF noted that GraphCast predicted the track with skill comparable to or exceeding operational NWP models at 5-day lead time. This was one of the first real-time tropical cyclone evaluations of a purely ML-based forecasting system during a significant event.

<60s
GraphCast 10-day global forecast time on single TPU
90%
ECMWF variables where GraphCast matched or exceeded performance at >3-day lead
40 yrs
ERA5 reanalysis data used for training (1979–2018)
1.8–5.6°C
CMIP6 equilibrium climate sensitivity range — still wide despite decades of modeling

Key Concepts

ERA5 ECMWF's fifth-generation atmospheric reanalysis — a comprehensive reconstruction of the global atmosphere from 1940 to present, blending observations with model output. The de facto training dataset for ML weather models.
Reanalysis A scientific method for developing comprehensive records of past atmospheric conditions by combining historical observations with modern numerical weather prediction models.
Distribution Shift When a model encounters inputs outside its training data. A critical concern for ML climate models: they've been trained on historical climate, but must generalize to future climates that may have no historical analogue.

Lesson 1 Quiz

Climate Models & Machine Learning — 4 questions
What architectural approach does GraphCast use to represent the atmosphere?
Correct. GraphCast uses a GNN where atmospheric grid points become nodes and edges capture spatial relationships — allowing irregular resolution and efficient message passing across scales.
Not quite. GraphCast uses a Graph Neural Network (GNN) architecture, treating atmospheric locations as nodes in a graph — not a CNN, LSTM, or random forest.
Why does halving a GCM's horizontal resolution require roughly eight times the compute?
Correct. Halving grid spacing quadruples the number of horizontal cells (2² = 4), and the Courant–Friedrichs–Lewy condition requires halving the timestep — a combined factor of roughly 8×.
Incorrect. The ~8× factor comes from the interaction of two effects: four times more horizontal grid cells (2² factor), plus a required halving of the model timestep for numerical stability (CFL condition).
Which dataset forms the primary training foundation for most ML weather forecasting systems including GraphCast and Pangu-Weather?
Correct. ERA5 — ECMWF's fifth-generation atmospheric reanalysis covering 1940 to present — is the de facto training dataset for ML weather models. Both GraphCast and Pangu-Weather were trained on ERA5.
Incorrect. ERA5 (ECMWF's fifth-generation reanalysis) is the primary training dataset. It provides 40+ years of comprehensive global atmospheric data at consistent resolution — ideal for supervised ML training.
What is the central concern about ML climate models facing "distribution shift"?
Correct. ML models trained on ERA5 (1940–2018 climate) may encounter atmospheric conditions in a warming world that have no precedent in their training data — making their generalization to novel climate states a genuine open question.
Distribution shift means encountering inputs outside training data. For ML climate models, the worry is that future warming produces atmospheric states with no historical analogue — so the models have never seen training examples resembling what they must forecast.

Lab 1 — ML Weather Model Analysis

Explore the trade-offs between data-driven and physics-based climate modeling

Your Mission

GraphCast and Pangu-Weather represent a genuine paradigm shift in atmospheric science — but they also carry significant limitations for climate (as opposed to weather) applications. In this lab, you'll interrogate those trade-offs with an AI assistant that specializes in ML-climate intersections.

Try asking: "What's the difference between using ML for weather forecasting vs. climate projection?" — or — "Why can't we just train GraphCast on future climate scenarios?"
Climate AI Assistant
Lab 1
Welcome to Lab 1. I'm here to help you think through the intersection of machine learning and climate science — specifically the capabilities and limits of data-driven models like GraphCast. Ask me anything about ML weather forecasting, reanalysis data, distribution shift, or how these systems compare to physics-based GCMs.
Lesson 2 · AI & Climate Science

Remote Sensing, Satellites, and AI Pattern Recognition

Petabytes of Earth observation data — AI is the only tool fast enough to read it.
How does machine learning turn raw satellite imagery into actionable climate intelligence?

Between 2020 and 2023, Global Fishing Watch and Google used convolutional neural networks trained on Sentinel-1 SAR imagery and AIS transponder data to track every large vessel on Earth's oceans. Their 2023 Nature paper revealed that 75% of the world's industrial fishing vessels were invisible to traditional monitoring — operating without AIS. The AI system detected over 300,000 previously untracked vessels. This same infrastructure now maps deforestation, illegal mining, and methane flares with comparable precision.

The Data Avalanche

The European Space Agency's Sentinel satellite constellation alone generates over 8 terabytes of imagery per day. NASA's Earth Observing System has accumulated more than 10 petabytes since 1999. NOAA's GOES-16 and GOES-18 geostationary satellites each transmit imagery at roughly 1 million pixels every 30 seconds. The bottleneck is no longer data collection — it is analysis.

Manual analysis of satellite imagery by trained scientists cannot keep pace. A single Landsat scene covers 185×185 km at 30-meter resolution — 38,000 square kilometers of data in one image. Monitoring global forest cover, ice extent, urban heat islands, and agricultural land use simultaneously requires automated methods.

Convolutional Neural Networks for Earth Observation

CNNs were initially developed for photographic image classification (ImageNet, 2012). Their application to multispectral satellite imagery followed rapidly. The key adaptation: satellite sensors capture 4–12 spectral bands beyond visible light, including near-infrared, shortwave infrared, and thermal infrared. CNNs can be trained to exploit all bands simultaneously.

Semantic segmentation assigns a land-cover class to every pixel — forest, water, urban, bare soil, cropland. Models like DeepLab and U-Net, adapted for remote sensing, achieve accuracies exceeding 90% on standard benchmarks. The Allen Institute's Dynamic World product (2022), built in collaboration with Google Earth Engine, produces near-real-time global land cover at 10-meter resolution using Sentinel-2 imagery and deep learning — updated continuously rather than annually.

Case Study — Amazon Deforestation Detection

Brazil's INPE (National Institute for Space Research) operates DETER, a real-time deforestation alert system using Landsat and CBERS imagery. Since 2019, AI-assisted analysis has reduced alert latency from weeks to days. Between August 2019 and July 2020, DETER detected 11,088 km² of Amazon deforestation alerts — an area larger than Jamaica. The speed advantage of ML analysis is directly tied to enforcement response time.

Methane Detection from Space

Methane (CH₄) is 80× more potent than CO₂ over a 20-year horizon. Detecting and attributing methane plumes from satellites requires analyzing shortwave infrared absorption signatures in hyperspectral data — a task where ML has made dramatic recent advances.

MethaneSAT (launched March 2024) by the Environmental Defense Fund uses onboard and ground-based ML processing to detect methane emissions from oil and gas operations globally at sub-field scale. Its predecessor data: Carbon Mapper and the EMIT instrument aboard the International Space Station, which in 2022–2023 identified over 50 "ultra-emitter" methane point sources in Central Asia, the Middle East, and the United States, each releasing thousands of kilograms per hour.

EMIT — A Specific Example

NASA's EMIT (Earth Surface Mineral Dust Source Investigation) instrument, installed on the ISS in July 2022, was originally designed to map mineral dust. By late 2022, researchers adapted its hyperspectral imaging capability to detect methane plumes. EMIT identified a "super-emitter" complex in Turkmenistan spanning multiple gas facilities — estimated at over 50,000 kg/hour. The detection required ML-based spectral unmixing algorithms to separate methane signal from surface reflectance.

Ice Sheet and Sea Ice Monitoring

The National Snow and Ice Data Center (NSIDC) has used passive microwave satellite data since 1979 to track Arctic sea ice extent. Traditional algorithms rely on fixed brightness temperature thresholds. Deep learning approaches — including recurrent networks that incorporate temporal sequences of imagery — now outperform threshold methods, particularly in the "marginal ice zone" where melt ponds and leads create ambiguous microwave signatures.

For the Greenland Ice Sheet, NASA's Operation IceBridge (2009–2019) and the CryoSat-2 radar altimeter have generated mass balance estimates requiring ML interpolation across the ice sheet's complex topography. Recent work using CNNs to map subglacial lake drainage events — detected as subtle surface deformation patterns in ICESat-2 altimetry data — has identified over 130 previously unknown subglacial lakes beneath Antarctica.

8 TB
Daily data output from ESA Sentinel constellation
75%
Industrial fishing vessels invisible to traditional AIS monitoring (Global Fishing Watch, 2023)
50,000+
kg/hour methane from single Turkmenistan complex detected by EMIT
10 m
Resolution of Dynamic World near-real-time global land cover product

Key Concepts

Semantic Segmentation Pixel-level classification of imagery into labeled categories. For satellite data this means assigning every pixel a land-cover type — enabling wall-to-wall mapping without manual delineation.
SAR (Synthetic Aperture Radar) Active radar imaging that works through clouds and at night. Sentinel-1 SAR data is critical for monitoring floods, deforestation, and vessel detection where optical imagery is obscured.
Hyperspectral Imaging Capturing hundreds of narrow wavelength bands across the electromagnetic spectrum. Each material has a unique spectral signature — enabling detection of gases like methane that absorb at specific wavelengths.

Lesson 2 Quiz

Remote Sensing & AI — 4 questions
What percentage of industrial fishing vessels were found to be invisible to traditional AIS monitoring in the 2023 Global Fishing Watch study?
Correct. The 2023 Nature paper from Global Fishing Watch found that 75% of industrial fishing vessels operated without AIS transponders — effectively invisible to conventional monitoring. AI analysis of Sentinel-1 SAR imagery revealed over 300,000 such vessels.
Incorrect. The 2023 Global Fishing Watch study found 75% of industrial fishing vessels were invisible to traditional AIS monitoring — far more than most observers expected. This finding drove significant policy discussion about maritime surveillance gaps.
Which NASA instrument, originally designed for mineral dust mapping, was repurposed to detect large methane point sources from the ISS?
Correct. EMIT was installed on the ISS in July 2022 for mineral dust mapping. Its hyperspectral imaging capability was quickly adapted to detect methane plumes — including a >50,000 kg/hour super-emitter complex in Turkmenistan.
Incorrect. EMIT (Earth Surface Mineral Dust Source Investigation), installed on the ISS in 2022, was the instrument repurposed for methane detection. Its hyperspectral capability allows it to distinguish methane absorption signatures from surface reflectance.
Why is SAR (Synthetic Aperture Radar) particularly valuable for environmental monitoring compared to optical satellite imagery?
Correct. SAR is an active radar system that emits its own microwave signal. This makes it cloud-penetrating and daylight-independent — critical for monitoring floods, deforestation, and maritime traffic in cloudy tropical regions.
SAR's key advantage is its active radar principle: it emits microwave energy and measures the return signal. This means clouds are transparent to it and it needs no sunlight — enabling year-round, all-weather monitoring impossible with passive optical sensors.
Google Earth Engine's "Dynamic World" product updates global land cover at what resolution, and how frequently?
Correct. Dynamic World uses Sentinel-2 imagery and deep learning to produce 10-meter resolution land cover classifications updated near-continuously — a dramatic improvement over annual static maps from traditional programs.
Dynamic World achieves 10-meter resolution using Sentinel-2 data and is updated near-continuously (not annually or monthly). This makes it among the most temporally and spatially detailed global land cover products ever created.

Lab 2 — Satellite Data & Climate Intelligence

Investigate how AI interprets Earth observation data for climate monitoring

Your Mission

From detecting illegal fishing to tracking methane super-emitters, AI-powered satellite analysis is transforming what we can monitor on Earth. In this lab, explore how different ML techniques apply to different remote sensing challenges — and what the limits of satellite-based monitoring are.

Try asking: "How does spectral unmixing help detect methane from space?" — or — "What makes the marginal ice zone hard to classify from satellite data?"
Remote Sensing AI Assistant
Lab 2
Welcome to Lab 2. I specialize in AI applications for satellite-based Earth observation — covering topics like CNNs for land cover mapping, SAR vessel detection, hyperspectral methane sensing, and ice sheet monitoring. What aspect of remote sensing AI would you like to explore?
Lesson 3 · AI & Climate Science

Extreme Event Attribution and Prediction

When a hurricane intensifies in six hours, or a heatwave kills thousands — AI is changing how fast we understand why.
How does machine learning shift extreme weather from something we explain after the fact to something we anticipate before it happens?

In late June 2021, the Pacific Northwest of North America experienced temperatures 40–50°F above normal. Portland, Oregon reached 116°F (46.6°C) — shattering its previous record by 4.4°F. Lytton, British Columbia hit 49.6°C (121.3°F), Canada's all-time national record. Within weeks, a rapid attribution study by World Weather Attribution found that the event was "virtually impossible without climate change." The analysis, using both physics-based models and statistical methods, was completed in under two weeks — a timeline that would have been impossible a decade earlier.

What Is Extreme Event Attribution?

Attribution science asks: did climate change make this event more likely, more severe, or both? The methodology typically involves running large ensembles of climate model simulations — hundreds to thousands of runs — comparing a world with observed greenhouse gas concentrations against a counterfactual world with pre-industrial concentrations.

Traditional attribution required months. The computational bottleneck was running enough model simulations to establish statistical significance. ML is changing this in two ways: first, by dramatically accelerating the simulations themselves; second, by enabling new statistical approaches that extract attribution signals from observational data alone.

World Weather Attribution Network

Established in 2014, WWA has conducted over 50 rapid attribution studies of major weather events. Their Pacific Northwest heatwave study (2021) concluded the event was at least 150 times more likely in the current climate than in pre-industrial conditions — and would have been 0.5°C cooler without anthropogenic warming. The analysis used CMIP6 model ensembles and observational station data.

Rapid Intensification in Tropical Cyclones

Rapid intensification (RI) — defined as a 35+ knot increase in maximum sustained winds in 24 hours — is one of the most dangerous and difficult-to-forecast phenomena in meteorology. Hurricane Michael (2018) intensified from Category 2 to Category 5 in the final 24 hours before making landfall near Mexico Beach, Florida. Hurricane Ian (2022) underwent RI over the Gulf of Mexico, making landfall at Category 4 with 150 mph winds.

NOAA's Hurricane Intensity Forecasting Improvements project has incorporated ML models alongside traditional statistical and dynamical models. The Statistical Hurricane Intensity Prediction Scheme (SHIPS) was the standard for decades. Deep learning approaches — particularly those incorporating sea surface temperature gradients, vertical wind shear profiles, and inner-core microwave imagery — have demonstrated skill improvements of 10–20% over SHIPS for RI prediction at 24-hour lead times in cross-validation studies.

Hurricane Lee — September 2023

Hurricane Lee underwent one of the most extreme rapid intensification events on record in the Atlantic basin, intensifying from Category 1 to Category 5 in approximately 24 hours on September 7, 2023 — a 100-knot intensification in 24 hours. NOAA's operational ML-augmented intensity forecasts correctly flagged high RI probability 12–18 hours in advance. The National Hurricane Center's official forecast, which incorporated these ML guidance products, issued RI advisories that gave residents along the eventual track additional preparation time.

Heatwave Prediction with Neural Networks

A landmark 2023 study in Nature Geoscience by Chattopadhyay et al. demonstrated that artificial neural networks trained on ERA5 data could predict the probability of extreme summer heat events in Europe and North America up to 15 days in advance — well beyond the deterministic predictability horizon for individual weather events (roughly 10 days). The ANNs learned to recognize large-scale atmospheric precursor patterns in the stratosphere and Pacific sea surface temperatures that traditional forecast models do not leverage for extended-range prediction.

The European heatwaves of 2003, 2019, and 2022 were all preceded by detectable precursor patterns in the stratospheric polar vortex and Rossby wave configurations. ML models trained to recognize these patterns can issue probabilistic long-range heat alerts — not specific temperature forecasts, but probability distributions that inform public health preparedness decisions weeks in advance.

Flood Forecasting: Google's Flood Hub

In 2018, Google began deploying AI-based flood forecasting models in India and Bangladesh — two of the world's most flood-affected nations. By 2023, Google's Flood Hub system covered river reaches in 80 countries representing over 460 million people at risk of riverine flooding. The system uses a combination of hydrological models, ML-based stage-discharge relationships, and real-time rainfall data to issue 7-day flood inundation forecasts.

During the 2023 monsoon season, Flood Hub issued alerts for multiple major flood events in India's Brahmaputra and Ganges basins. Google reported that forecasts were reaching end users — via Android alerts in affected areas — with 24–72 hour lead times for events that previously had warning windows of hours.

150×
Increased likelihood of 2021 Pacific Northwest heatwave in current vs. pre-industrial climate (WWA)
15 days
Advance warning of extreme heat probability achievable with trained ANNs (Chattopadhyay et al., 2023)
460M
People covered by Google Flood Hub's AI forecasting system (2023)
100 kt
Hurricane Lee's intensification in 24 hours — September 2023

Key Concepts

Rapid Intensification (RI) A tropical cyclone strengthening by ≥35 knots in 24 hours. Notoriously difficult to forecast due to complex ocean-atmosphere interactions. ML models incorporating SST gradients and microwave inner-core structure have improved RI skill.
Attribution Science Quantitative methods for determining the extent to which climate change altered the probability or intensity of a specific extreme weather event. Rapid attribution now uses ML-accelerated ensembles to deliver results in days rather than months.
Rossby Waves Large-scale meanders in the upper-level westerly winds that mediate heat transport between polar and tropical regions. AI models have learned to identify Rossby wave configurations that precede extreme heat events weeks in advance.

Lesson 3 Quiz

Extreme Events & AI — 4 questions
The World Weather Attribution study of the 2021 Pacific Northwest heatwave concluded the event was how many times more likely in the current climate vs. pre-industrial conditions?
Correct. WWA's rapid attribution study found the Pacific Northwest heatwave of June 2021 was at least 150 times more likely in today's climate — and would have been approximately 0.5°C cooler without anthropogenic greenhouse gas forcing.
The WWA study found the event was at least 150 times more likely in current climate — a striking result reflecting both the extreme magnitude of the event and the strong climate change signal on summertime heat in the Pacific Northwest.
What is the meteorological definition of "rapid intensification" for tropical cyclones?
Correct. NHC defines rapid intensification as a ≥35-knot increase in maximum sustained winds over 24 hours. This threshold marks events that are particularly dangerous because they outpace official forecast uncertainty cones and public preparation timelines.
The operational NHC definition of rapid intensification is ≥35 knots of intensification in 24 hours. This is the threshold used in NOAA verification statistics and the benchmark against which ML intensity forecasting improvements are measured.
How far in advance can trained artificial neural networks predict the probability of extreme summer heat events, according to the 2023 Nature Geoscience study?
Correct. The Chattopadhyay et al. 2023 study showed ANNs trained on ERA5 could predict extreme heat probability up to 15 days ahead by recognizing stratospheric and Pacific SST precursor patterns that traditional models don't use for extended-range forecasting.
The 2023 Nature Geoscience study (Chattopadhyay et al.) demonstrated 15-day advance probabilistic prediction of extreme heat — exceeding the roughly 10-day deterministic predictability horizon — by exploiting stratospheric and large-scale SST precursor signals.
As of 2023, approximately how many people were covered by Google's AI-based Flood Hub forecasting system?
Correct. By 2023, Google's Flood Hub covered river reaches in 80 countries representing over 460 million people at risk of riverine flooding — delivering 7-day inundation forecasts with lead times of 24–72 hours for major events.
By 2023, Google Flood Hub covered 80 countries and over 460 million people — one of the largest-scale deployments of AI for climate-related disaster risk reduction. Alerts are delivered directly via Android notifications in affected regions.

Lab 3 — Extreme Event AI Workshop

Interrogate how AI predicts, attributes, and communicates climate extremes

Your Mission

Extreme weather events — heatwaves, hurricanes, floods — are where AI's climate capabilities most directly affect human lives. In this lab, explore the technical and societal dimensions of AI-based extreme event forecasting and attribution.

Try asking: "What atmospheric patterns do ANNs use to predict European heatwaves two weeks out?" — or — "How does attribution science distinguish climate change from natural variability in a specific heatwave?"
Extreme Weather AI Assistant
Lab 3
Welcome to Lab 3. I'm your guide for AI in extreme weather forecasting and attribution. Ask me about rapid intensification prediction, heatwave precursor patterns, flood forecasting systems, or the methodology behind World Weather Attribution studies.
Lesson 4 · AI & Climate Science

AI for Carbon Monitoring and Climate Solutions

From measuring emissions to designing fusion reactors — AI as climate tool and climate challenge.
Can the same technology that accelerated fossil fuel exploration become the engine of decarbonization?

On December 5, 2022, researchers at the National Ignition Facility in Livermore, California achieved nuclear fusion ignition for the first time in history — 3.15 megajoules of energy released from 2.05 megajoules of laser energy delivered. The shot design, target geometry, and laser pulse timing that made it possible relied on machine learning optimization of fusion plasma physics models. This was not AI solving fusion, but it illustrated the accelerating role of ML in complex physical system design.

Carbon Measurement, Reporting, and Verification

The Paris Agreement depends on accurate greenhouse gas accounting. Currently, national inventories rely heavily on "bottom-up" estimates derived from economic activity data and emission factors. These estimates can be verified — or contradicted — by "top-down" atmospheric inversion methods that work backward from observed atmospheric CO₂ and CH₄ concentrations to infer surface fluxes.

ML is enhancing both approaches. For bottom-up estimation, Google's Environmental Insights Explorer uses ML to analyze building permits, satellite imagery, and street-level imagery to estimate city-level CO₂ emissions without requiring self-reported data. For top-down verification, the CarbonTracker system and its successors use neural networks to improve the inversion algorithms that translate atmospheric measurements into surface flux maps.

Case Study — Global Carbon Project 2023

The Global Carbon Project's 2023 annual report used ML-enhanced ocean carbon flux estimates (SOCOM-MPI neural network ensemble) to refine ocean CO₂ uptake estimates. The ocean absorbs roughly 26% of human CO₂ emissions annually — but sparse ship-based measurements make direct estimation uncertain. ML interpolation of sparse surface ocean pCO₂ measurements reduced the uncertainty in global ocean carbon uptake by approximately 30% compared to previous interpolation methods.

Decarbonizing the Power Grid

In 2019, DeepMind announced that reinforcement learning had improved the energy efficiency of Google's data center cooling systems by 40% — reducing cooling energy (roughly 40% of data center energy) by that amount. The system learned to anticipate cooling demands rather than react to them, optimizing chiller configurations, cooling tower fan speeds, and pump rates.

The same year, DeepMind partnered with Ørsted to apply ML to wind farm power output prediction. By predicting wind power output 36 hours ahead, operators could make firm capacity commitments to the grid — increasing the value of wind energy by roughly 20% in grid markets that reward predictability. Without accurate forecasting, intermittent renewables must be discounted because they cannot make firm delivery commitments.

AlphaFold's Climate Parallel

DeepMind's AlphaFold protein structure predictor (2020) demonstrated that ML could solve a 50-year grand challenge in biology. Researchers now apply similar approaches to materials science for climate: AI-accelerated screening of solid-state electrolytes for grid-scale batteries (Microsoft's MatterGen, 2024), catalysts for green hydrogen production, and perovskite photovoltaic stability. These are not demonstrations — materials identified by ML are entering experimental validation pipelines.

The Energy Cost of AI Itself

A critical tension in AI-for-climate work is that training large AI models is itself energy-intensive. A 2019 study by Emma Strubell et al. estimated that training a large Transformer NLP model produced roughly 284 tonnes of CO₂-equivalent — comparable to five round-trip transatlantic flights per parameter training run. GPT-4-scale models represent orders of magnitude more compute.

However, the picture is nuanced. Microsoft, Google, and AWS — the three dominant cloud providers running most AI workloads — have made significant investments in matching their electricity consumption with renewable energy certificates. Google reported operating carbon-free 24/7 on an hourly basis in 64% of its global data center energy in 2022. NVIDIA's H100 GPUs are roughly 6× more energy-efficient per training FLOP than the V100 generation.

The lifecycle assessment question remains: do the climate benefits of AI applications (improved renewable forecasting, materials discovery, demand optimization) exceed the emissions from training and running those systems? Current evidence suggests they do, for well-targeted applications — but the rapid scaling of AI compute makes this calculation dynamic.

Nuclear Fusion and Materials Discovery

In February 2022, Google DeepMind published work in Nature demonstrating that reinforcement learning could control the plasma configuration in a tokamak fusion reactor at the Swiss Federal Institute of Technology (EPFL). The RL system controlled 19 magnetic coils in real time to maintain stable plasma shapes — including novel configurations that human operators had not previously attempted. This represented the first time deep RL had controlled a physical nuclear fusion device.

DeepMind followed this with work on GNoME (Graph Networks for Materials Exploration) in November 2023, identifying 2.2 million new stable crystal structures — including 380,000 predicted to be stable enough for experimental synthesis. Among the candidates are potential new battery materials, solar absorbers, and high-temperature superconductors relevant to grid-scale energy storage and transmission.

40%
Data center cooling energy reduction from DeepMind RL optimization (2019)
2.2M
New stable crystal structures identified by DeepMind GNoME (2023)
30%
Reduction in ocean CO₂ uptake uncertainty from ML interpolation (Global Carbon Project 2023)
20%
Increase in wind energy market value from 36-hour ML forecasting (Ørsted/DeepMind)

Key Concepts

Atmospheric Inversion A top-down method that uses observed atmospheric greenhouse gas concentrations and transport models to infer surface emission sources and sinks. ML improves inversion algorithms, tightening constraints on national emission estimates.
Reinforcement Learning (RL) A training paradigm where an agent learns to take actions in an environment to maximize cumulative reward. Used by DeepMind for data center cooling and fusion plasma control — both sequential decision problems with complex feedback loops.
Life Cycle Assessment (LCA) A methodology for quantifying the environmental impacts of a product or process from cradle to grave. Applying LCA to AI systems asks: does training this model cause more emissions than it saves through its downstream application?

Lesson 4 Quiz

Carbon Monitoring & Climate Solutions — 4 questions
DeepMind's reinforcement learning system reduced Google data center cooling energy by approximately how much?
Correct. DeepMind's RL system, deployed in 2019, reduced cooling energy by 40% by learning to anticipate thermal loads rather than react to them — adjusting chiller configurations, cooling tower fans, and pump rates in real time.
DeepMind's reinforcement learning system achieved a 40% reduction in data center cooling energy — a significant result since cooling typically represents ~40% of total data center energy consumption.
What was significant about DeepMind's February 2022 Nature paper on tokamak plasma control?
Correct. The DeepMind/EPFL work published in Nature (2022) was the first demonstration of deep reinforcement learning controlling magnetic coils in a real tokamak to maintain plasma configurations — including novel shapes not previously attempted by human operators.
The key milestone was first-ever deep RL control of plasma in a physical fusion device (EPFL's TCV tokamak). Net energy gain came separately at the NIF in December 2022 via laser inertial confinement — a different fusion approach.
The Global Carbon Project's ML-enhanced ocean carbon flux estimates (SOCOM-MPI) reduced uncertainty in global ocean CO₂ uptake by approximately how much?
Correct. The SOCOM-MPI neural network ensemble, used in the GCP 2023 report, reduced ocean CO₂ uptake uncertainty by approximately 30% compared to previous interpolation methods — critical given the ocean absorbs about 26% of annual human CO₂ emissions.
The ML-enhanced SOCOM-MPI ensemble achieved approximately a 30% reduction in uncertainty for global ocean CO₂ uptake estimates. This matters enormously since the ocean absorbs ~26% of human CO₂ emissions annually — a flux that must be accurately known for global carbon budgeting.
How many new stable crystal structures did DeepMind's GNoME system identify in its November 2023 publication?
Correct. GNoME (Graph Networks for Materials Exploration) identified 2.2 million new stable crystal structures, of which 380,000 were predicted stable enough for experimental synthesis. These include potential battery materials, solar absorbers, and superconductors.
GNoME identified 2.2 million new stable crystal structures total. The 380,000 figure is the subset predicted stable enough for experimental synthesis — both numbers represent dramatic expansions of the known materials space for clean energy applications.

Lab 4 — AI, Carbon & Clean Energy

Explore the dual role of AI as both a climate tool and a climate challenge

Your Mission

AI is simultaneously a tool for climate solutions and a source of energy demand. In this lab, explore the applications of AI in carbon monitoring, materials discovery, grid optimization, and nuclear fusion — and interrogate the critical question of whether AI's climate benefits outweigh its energy costs.

Try asking: "How does reinforcement learning optimize wind farm output?" — or — "What's the carbon footprint of training a large language model and how does it compare to its climate benefits?"
Climate Solutions AI Assistant
Lab 4
Welcome to Lab 4. I'm here to explore AI's role in climate solutions — and its own climate footprint. Ask me about carbon monitoring, materials discovery with GNoME, fusion plasma control, grid optimization, or the energy cost vs. climate benefit trade-offs of large-scale AI deployment.

Module Test — AI for Climate Science

15 questions · Score 80% or higher to pass
1. GraphCast produces a 10-day global weather forecast in under 60 seconds. What hardware does it run on?
Correct. GraphCast runs on a single Google TPU chip — a dramatic demonstration of efficiency compared to the supercomputer clusters required by physics-based NWP systems.
GraphCast runs on a single TPU chip — the key point being that a single specialized accelerator chip now outperforms supercomputer-scale physics-based models on many metrics.
2. In CMIP6 models, the equilibrium climate sensitivity (ECS) range spans from 1.8°C to 5.6°C per CO₂ doubling. What is the primary driver of this uncertainty?
Correct. Clouds — and disagreements in how different GCMs parameterize cloud feedbacks — account for roughly half the spread in ECS estimates. This has been known since the 1979 Charney Report.
Cloud parameterization is the dominant source of ECS uncertainty. Different models produce very different cloud feedbacks under warming — some clouds amplify warming, others moderate it — and models disagree substantially on which effect dominates.
3. What does "emulation" mean in the context of ML for climate modeling?
Correct. An ML emulator learns the input-output behavior of an expensive physical model component — like mesoscale ocean eddy parameterizations — and reproduces it 100× faster without running the original physics code.
Emulation means training an ML model to replicate the behavior of an expensive physics-based component. Columbia's M2LInES project demonstrated this for ocean eddy parameterizations, achieving 100× speedup.
4. ESA's Sentinel satellite constellation generates approximately how much data per day?
Correct. Sentinel generates over 8 TB per day — making it impossible for manual analysts to process in near-real-time and driving the need for automated ML-based analysis pipelines.
ESA's Sentinel constellation generates over 8 terabytes of imagery daily — a volume that fundamentally requires automated AI analysis to be useful for near-real-time monitoring applications.
5. NASA's EMIT instrument was originally designed to map what, before being repurposed for greenhouse gas detection?
Correct. EMIT (Earth Surface Mineral Dust Source Investigation) was designed to map mineral dust composition. Its hyperspectral imaging capability was quickly adapted to detect methane absorption signatures.
EMIT stands for Earth Surface Mineral Dust Source Investigation — it was designed to map where airborne dust originates. Its hyperspectral imaging proved equally capable of detecting methane plumes from orbit.
6. Why is SAR imagery particularly valuable for monitoring tropical deforestation?
Correct. Tropical regions are often cloud-covered, making optical satellite monitoring unreliable. SAR's microwave signals penetrate clouds, enabling continuous monitoring regardless of cloud cover — critical for detecting deforestation events quickly.
The key advantage is cloud penetration. Tropical forests are frequently cloud-covered — sometimes for weeks — making optical monitoring unreliable. SAR microwaves pass through clouds, enabling consistent repeat-pass monitoring.
7. The 2023 Nature paper from Global Fishing Watch used AI to detect how many previously untracked industrial fishing vessels?
Correct. The AI system combining CNNs with Sentinel-1 SAR imagery detected over 300,000 previously untracked industrial fishing vessels — representing 75% of the global industrial fleet operating without AIS transponders.
The Global Fishing Watch AI detected over 300,000 previously invisible vessels — representing 75% of the industrial fleet. This revealed massive gaps in existing maritime monitoring infrastructure.
8. Rapid intensification in a tropical cyclone is formally defined as how much wind speed increase in 24 hours?
Correct. NHC defines rapid intensification as ≥35 knots of maximum sustained wind increase in 24 hours. ML models incorporating sea surface temperature gradients and inner-core microwave imagery have improved RI prediction by 10–20% over traditional statistical schemes.
35 knots per 24 hours is the operational RI threshold used by the National Hurricane Center. Hurricane Lee (2023) exceeded this dramatically, intensifying by ~100 knots in 24 hours.
9. The World Weather Attribution study of the 2021 Pacific Northwest heatwave found that the event would have been how much cooler without anthropogenic climate change?
Correct. WWA found the event would have been ~0.5°C cooler without climate change — but the event's extreme deviation from historical norms made it at least 150× more likely in current climate. Both the magnitude and frequency aspects of attribution matter.
WWA found the event would have been ~0.5°C cooler without anthropogenic forcing. However, this modest-sounding temperature difference translates to a massive increase in probability (150×) given the extreme tails of the temperature distribution.
10. Neural networks trained on ERA5 data demonstrated the ability to predict extreme summer heat probability at what lead time (Chattopadhyay et al., Nature Geoscience 2023)?
Correct. The ANNs achieved 15-day probabilistic prediction of extreme heat — beyond the ~10-day deterministic predictability limit — by exploiting stratospheric polar vortex conditions and Pacific SST patterns as precursors.
15 days was the key result. This exceeded the conventional ~10-day predictability horizon by exploiting large-scale atmospheric precursors in the stratosphere and Pacific — patterns that traditional medium-range forecasting doesn't leverage for heat prediction.
11. Google's Flood Hub AI forecasting system, as of 2023, serves populations at flood risk in how many countries?
Correct. By 2023, Flood Hub covered 80 countries representing over 460 million people at riverine flood risk, with 7-day forecasts issued via Android alerts to users in affected areas.
Google's Flood Hub expanded to 80 countries by 2023, covering over 460 million people. Starting from India and Bangladesh in 2018, the system's geographic expansion was enabled by the scalability of its ML-based approach.
12. DeepMind's GNoME system identified 2.2 million new stable crystal structures. Of those, how many were predicted stable enough for experimental synthesis?
Correct. Of the 2.2 million new stable crystal structures, approximately 380,000 were predicted stable enough for experimental synthesis — candidates for new battery materials, solar absorbers, and high-temperature superconductors.
380,000 of the 2.2 million structures were flagged as experimentally synthesizable candidates. GNoME identified more new stable materials in one publication than had been discovered by all experimental methods in the history of materials science.
13. What technique does the SOCOM-MPI system use to reduce uncertainty in estimates of the ocean's carbon uptake?
Correct. SOCOM-MPI uses a neural network ensemble to interpolate sparse ship-based surface ocean pCO₂ measurements to global coverage — reducing uncertainty in the ocean CO₂ sink by ~30%.
SOCOM-MPI interpolates sparse pCO₂ measurements across the global ocean using a neural network ensemble. Ship-based pCO₂ measurements are geographically concentrated in shipping lanes, leaving vast ocean areas unmeasured — ML interpolation fills these gaps.
14. DeepMind's 2022 Nature paper on fusion plasma control was conducted at which facility?
Correct. The RL plasma control system was demonstrated on EPFL's TCV (Tokamak à Configuration Variable) tokamak at the Swiss Federal Institute of Technology in Lausanne — the first deep RL system to control plasma in a real fusion device.
The DeepMind/Nature 2022 work used EPFL's TCV tokamak in Switzerland. NIF achieved fusion ignition (December 2022) via laser inertial confinement — a separate facility and fusion approach.
15. Regarding AI's own energy footprint, what is the most accurate characterization of the current situation?
Correct. Training large models produces significant CO₂ (Strubell et al. estimated ~284 tonnes for large NLP models in 2019), but major cloud providers are increasing renewable energy matching and hardware efficiency is improving. Net climate benefit depends on application specificity and is an active research question.
The nuanced answer is correct: AI training is genuinely energy-intensive, but for well-targeted climate applications (flood forecasting, wind prediction, materials discovery) the downstream climate benefits are likely to exceed training emissions. However, rapid scaling of AI compute makes this calculation dynamic and uncertain.