AI for climate research: weather forecasts, climate models and their limits

How AI weather forecasting works

Traditional forecasting is numerical weather prediction (NWP). A supercomputer takes an estimate of the atmosphere's current state and solves the equations of fluid motion, heat and moisture forward in time, step by step. A machine learning weather model does not solve those equations. It learns from decades of past weather how one state of the atmosphere turns into the next, then applies what it learned to today's state.

Google DeepMind's GraphCast is the reference example. It was trained on 39 years (1979 to 2017) of ECMWF's ERA5 reanalysis, according to the paper's preprint. It takes two inputs, the state of the weather six hours ago and the state now, and predicts the state six hours ahead. Feeding each prediction back in as the next input rolls the forecast out to 10 days. DeepMind describes the model as a graph neural network on a 0.25° grid (about 28 km at the equator, more than a million points) that predicts five surface variables and six atmospheric variables at each of 37 altitude levels.

The results changed what the field expected. In the peer-reviewed version, published in Science on 14 November 2023, GraphCast outperformed the most accurate operational deterministic systems on 90% of 1,380 verification targets, and it produces a 10-day global forecast in under a minute on a single Google TPU v4. DeepMind compares that with hours on a supercomputer with hundreds of machines for ECMWF's High Resolution Forecast (HRES). The cost moved to training, which took roughly four weeks on 32 Cloud TPU v4 devices.

From one forecast to a range of outcomes

A single forecast hides its uncertainty, and beyond a few days the uncertainty is much of the information. GenCast, published in Nature on 4 December 2024, is a diffusion model that generates an ensemble of realistic 15-day forecasts at 0.25°. It had greater skill than ECMWF's physics-based ensemble (ENS) on 97.2% of 1,320 targets, and each 15-day forecast takes about 8 minutes on one Cloud TPU v5, with the members generated in parallel.

Google's WeatherNext 2, announced on 17 November 2025, generates hundreds of possible scenarios from a single starting point, each in under a minute on a single TPU, and surpasses Google's previous WeatherNext model on 99.9% of variables and lead times from 0 to 15 days. Other labs have followed. Microsoft's Aurora, published in Nature in May 2025, is a foundation model trained on more than a million hours of geophysical data and fine-tuned to outperform operational forecasts of air quality, ocean waves, tropical cyclone tracks and high-resolution weather.

Which AI weather models forecasters use today

The clearest sign of maturity is that forecast agencies now run these models themselves, beside their physics models rather than in place of them. ECMWF made AIFS Single and then AIFS ENS operational in 2025, and NOAA launched three AI models as one suite that December:

ECMWF calls its AI and physics systems complementary. Its physics-based ensemble runs at 9 km, and ECMWF says it is still needed for high-resolution detail and for processes that couple the atmosphere to the rest of the Earth system, which is why the centre is also researching hybrids of the two. NOAA's suite started from GraphCast, fine-tuned on analyses from NOAA's own Global Data Assimilation System.

Hurricane forecasting shows how agencies adopt these models in practice. Google DeepMind's experimental cyclone model, launched on its Weather Lab site in June 2025, predicts a cyclone's formation, track, intensity, size and shape as 50 scenarios up to 15 days ahead, and forecasters at the US National Hurricane Center (NHC) saw its live predictions during the 2025 season. In a Q&A published by the National Weather Service, NHC's science operations officer says forecasters began using AI models as guidance during that season, alongside physics models, and that the official forecast, which weighs all of them, still outperforms any individual model.

What AI weather models still depend on

Every model above learns from, and starts from, products of the physics-based system it is compared against. That dependency is easy to miss in headlines about AI replacing weather models.

Training data comes from reanalysis. A reanalysis combines past observations with a forecast model to produce a consistent, gap-free record. ERA5, produced by ECMWF for the EU's Copernicus Climate Change Service, provides hourly data on a 0.25° grid with 37 pressure levels from 1940 onward, updated daily about five days behind real time. ECMWF says ERA5 provides the training data for most of today's leading machine learning forecasting systems.

Each forecast starts from a physics-based analysis. AIFS uses the same initial conditions as ECMWF's physics model: a previous short-range forecast combined with around 60 million quality-controlled observations from satellites, aircraft, ships, buoys and ground stations. For the ensemble version, ECMWF states that AIFS ENS relies on physics-based data assimilation to generate its initial conditions.

Research is working on both dependencies. Aardvark Weather, published in Nature in March 2025, ingests observations directly and produces global and local station forecasts without running an NWP system at deployment time, although its modules can be pretrained on ERA5 and its global forecasts were scored against it. In July 2026 ECMWF showed a prototype reanalysis from AIFS-DOP, a version of AIFS trained on observations alone. It rebuilt 1981 to 2022 on a roughly 112 km grid in a single working day and captured the large-scale circulation, but underestimated hurricane intensity, as expected at that grid spacing. ECMWF says operational reanalyses such as ERA5, and ERA6 now in production, will remain the trusted foundation and the training data for AI forecasting.

Climate emulators: many more simulated years for the same compute

Weather and climate are different problems. A weather forecast asks what the atmosphere will do over the next two weeks from today's state. A climate simulation asks how the statistics of weather change under different conditions, such as sea surface temperatures and greenhouse gas concentrations, and it must stay stable for decades while responding correctly to those conditions. A climate emulator is a fast learned model that imitates that behaviour.

NeuralGCM, published in Nature on 22 July 2024, is a hybrid. A differentiable solver handles the large-scale dynamics of the atmosphere, and neural networks represent small-scale processes such as cloud formation. It is competitive with ECMWF's ensemble for 1 to 15-day forecasts, and with prescribed sea surface temperatures it tracks climate metrics for multiple decades. Google Research reports that its 1.4° model is more than 3,500 times faster than X-SHiELD, a high-resolution atmospheric model: a simulated year takes 8 minutes on a single TPU instead of 20 days on a supercomputer with 13,000 CPUs. It is written in JAX, one of the libraries in our comparison of AI frameworks.

ACE2, the Ai2 Climate Emulator (npj Climate and Atmospheric Science, May 2025), takes the pure machine learning route. It is a 450-million-parameter model with 1° resolution, 6-hour steps and eight vertical layers that exactly conserves global dry air mass and moisture and simulates about 1,500 years per wall-clock day. It reproduces 80 years of atmospheric variability, including the response to El Niño and the global temperature trend.

The value is volume. When a simulated year costs minutes rather than weeks, researchers can run much larger ensembles and more scenarios for the same budget. Both papers also state the catch plainly: NeuralGCM does not extrapolate to substantially different future climates, and ACE2's sensitivities to separately changing sea surface temperature and carbon dioxide are not entirely realistic. That is why, as of September 2026, projections of a warmer future still start from physics-based Earth system models, like the CMIP6 ensemble used in the downscaling work below.

Downscaling: from global grids to local risk

Global climate models work on coarse grids, while agriculture, hydrology and natural hazard risk assessment need detail below about 10 km. Google researchers illustrate the gap in PNAS: a computing budget sufficient to simulate the global climate at 100 km resolution would fall short of what it takes to downscale a region the size of Spain to 10 km. The two established options each have a weakness. Dynamical downscaling runs a regional climate model driven by the global one; it is accurate but too expensive for large ensembles. Statistical downscaling is cheap but relies on relationships fitted to past data.

The PNAS paper (April 2025) combines physics and generative AI. A regional climate model first downscales the global model's output to an intermediate grid, then a diffusion model refines it to the target resolution (100 km to 45 km to 9 km in the paper's illustration). Tested against dynamically downscaled projections from the CMIP6 ensemble, it gave more accurate uncertainty bounds than dynamically downscaling a smaller ensemble, lower errors than popular statistical methods, and more faithful tails and correlations between variables.

The same idea works for weather. CorrDiff, published in February 2025, learned to turn 25 km reanalysis into the 2 km output of an operational regional weather model over Taiwan, using a deterministic model for the mean and a diffusion model for the residual detail. It learned from three years of data, sharpened cold fronts and intensified typhoons in case studies, and on a single GPU ran at least 22 times faster and 1,300 times more energy-efficiently than the numerical model that produced its training data. Its authors note that calibrating the model's uncertainty remains challenging.

Extreme events: earlier warnings and faster attribution

Extremes are where forecasts matter most, and where AI results deserve the closest reading.

Storms. DeepMind reported that a simple cyclone tracker run on GraphCast forecasts predicted cyclone movement more accurately than HRES, and that in September 2023 the live GraphCast model on ECMWF's website predicted about nine days ahead that Hurricane Lee would make landfall in Nova Scotia, while traditional forecasts settled on Nova Scotia about six days ahead. Track and intensity are different problems, though: NOAA's first AI model improved tracks but degraded intensity forecasts.

Floods. A Google model described in Nature in March 2024 forecasts river floods in ungauged watersheds, where no local gauge record exists to calibrate against. At up to five days' lead time, its reliability was similar to or better than the same-day nowcasts of GloFAS, the Copernicus Emergency Management Service's global flood awareness system, and it feeds free public forecasts in over 80 countries.

Attribution. Attribution asks a different question: how much did climate change alter this particular event? World Weather Attribution answers within days or weeks using observations and computer modelling. Its study of the 2021 heat wave in the Pacific Northwest of the US and Canada found that climate change made it at least 150 times more likely and 2 °C more intense.

One family of methods, storyline attribution, re-creates a specific event in climates with and without human influence, which normally takes expensive extra climate model runs. A 2024 study in Science Advances replaced those runs with convolutional neural networks trained on climate model simulations of 1850 to 2100. Given the calendar day, global mean temperature and daily maps of sea-level pressure, soil moisture and geopotential height, the networks predict daily maximum temperature; lowering the global temperature input produces a counterfactual version of the same weather. The authors estimate that the 2023 heat event in south-central North America was 1.18 to 1.42 °C warmer because of global warming, broadly in line with established techniques, and conclude that machine learning can make attribution rapid and low-cost.

Emissions monitoring: finding methane leaks from space

Methane is responsible for roughly 30% of the warming since pre-industrial times, and a small number of large point sources emit a disproportionate share, as the authors of MARS-S2L note. Finding those sources quickly is a detection problem, which is where machine learning fits.

UNEP's Methane Alert and Response System (MARS), launched at COP27 in 2022, is the first public global satellite system that detects very large methane emissions and notifies governments and companies about them. It combines data from more than 30 satellite instruments with AI models, and in 2026 it expanded from oil and gas to the coal and waste sectors.

The mechanism works like this. Methane absorbs sunlight in parts of the shortwave infrared, so a satellite that measures reflected sunlight in those bands can map extra methane above a site. Imaging spectrometers such as EMIT, PRISMA and EnMAP see it in fine spectral detail, but the standard retrieval, a matched filter, produces many false detections that a person must check, according to the team behind MARS's detection models. Multispectral satellites such as Sentinel-2 and Landsat measure only a few broad bands, but offer 20 to 30 m pixels, public global coverage every two days and, according to the MARS-S2L authors, detection of emissions above about 1 tonne per hour.

Two preprints from the MARS team report the results. For imaging spectrometers, combining several deep learning models cut false detections by over 74%; in 11 months of operation the system processed more than 25,000 hyperspectral products, helped verify 2,851 distinct methane leaks and led to 834 notifications. MARS-S2L, trained on more than 80,000 manually curated images, found 78% of plumes with an 8% false positive rate at 697 sites it had never seen, and has issued 2,776 notifications in 25 countries. Six persistent emitters were verifiably stopped, among them a super-emitter in Algeria that had released about 27,000 tonnes of methane a year for at least a decade. Both papers are preprints on arXiv, the first marked as under review, so treat the figures as the team's own until peer review.

The energy footprint of AI itself

AI's energy use cuts both ways, and the two scales should not be mixed up.

Per forecast, AI is far cheaper. ECMWF reports that AIFS uses about 1,000 times less energy to make a forecast than its physics model, and CorrDiff was 1,300 times more energy-efficient than the regional model it learned from. Training is the expensive part (GraphCast took about four weeks on 32 TPUs), but the trained model then runs every forecast after that.

Across all of AI, demand is rising fast. The IEA's Energy and AI report (10 April 2025) estimates that data centers used about 415 TWh in 2024, around 1.5% of the world's electricity, and projects that to more than double to about 945 TWh by 2030, with AI the most important driver. Emissions from data center electricity grow from 180 million tonnes today to 300 million tonnes by 2035 in its Base Case, and up to 500 million in its Lift-Off Case, still below 1.5% of energy sector emissions. The IEA's Electricity 2026 report (February 2026) still expects data centers to drive around half of US electricity demand growth through 2030.

The IEA's overall judgment is balanced: fears that AI will speed up climate change look overstated, and so do hopes that AI will fix it on its own. Wide adoption of existing AI applications could cut far more emissions than data centers produce, yet far less than climate goals require. For research teams, the practical habit is to report compute and energy next to skill, so savings per forecast can be weighed against the cost of training and of ever-larger ensembles. Why frontier AI keeps needing more compute is covered in our post on AI supercomputers and the path to AGI.

The limits: past data, explainability and evaluation

Training data comes from a climate that is changing

A model learns the weather it was shown. A 2025 PNAS study trained versions of the FourCastNet model on ERA5 from 1979 to 2015, some with every Category 3 to 5 tropical cyclone removed, and tested them on Category 5 storms from 2018 to 2023. All versions performed alike on global weather, but the one that had never seen a Category 3 to 5 storm could not forecast Category 5 ones: it did not extrapolate from weaker storms. Removing strong storms from only one ocean basin did less harm, which suggests the model can borrow from other basins. The authors expect the finding to apply to other current AI weather and climate models, and call such unseen extremes gray swans. In a warming climate, record-breaking events are exactly the ones missing from the record, the same reason NeuralGCM's authors warn against using it for substantially different future climates.

Explainability: knowing why, not only what

A physics model's forecast can be traced to equations and parameterizations a scientist can inspect. A neural network's cannot, at least not directly. NHC's science operations officer names this as a main drawback: AI systems are evolving quickly, and forecasters constantly need to understand why a model produced its result. Research is under way to show forecasters the reason as well as the answer. Some studies build interpretation into the method: the attribution study above used partial dependence analysis, which shows how the prediction moves as one input, global temperature, changes. Until explanations mature, agencies keep experts in the loop and treat AI output as one input among many.

Evaluation: better at what, and against what

Headline percentages need context. WeatherBench 2, the open benchmark many labs report against, warns that its headline scores capture important aspects of a good forecast, but not all of them. Before trusting an AI weather or climate result, check:

1. The baseline. GraphCast was compared with HRES, GenCast with ENS. A claim with no named physics baseline is a demonstration, not evidence.
2. The truth it was scored against. WeatherBench 2 evaluates operational models against ECMWF's operational analysis and machine learning models against ERA5. GraphCast's authors call that domain gap an important caveat to their own conclusions.
3. Held-out years. Test data must come after the training period, as in GraphCast's evaluation on 2018 onward.
4. The variables that matter to you. An average score can hide weak hurricane intensity (NOAA's AIGFS) or smoothed fields: training to minimize average error led GraphCast to produce blurrier forecasts at longer lead times.
5. Uncertainty. A single forecast cannot say how confident it is. Ensembles can, but their spread must be calibrated, which CorrDiff's authors still describe as challenging.
6. Extremes outside the record. Ask whether the model was tested on events stronger than anything in its training data.

The same discipline applies well beyond climate science. Any model, whether it forecasts storms or reads invoices, should be judged on held-out cases against a named baseline before anyone relies on it. If you are weighing AI for a workflow in your own organization, our AI and automation service starts from that rule: an evaluation set built from your real cases, passed before any model, prompt or source change goes live.