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