Machine Learning for Climate Prediction: What AI Can Forecast

A neural network called GraphCast, first reported in late 2022, can produce a global 10-day weather forecast at 0.25-degree resolution in under one minute, and its authors found it beat the most accurate operational deterministic systems on 90 percent of 1,380 verification targets. By 2025 the ECMWF, Europe's medium-range forecasting center, ran a machine learning system operationally. That describes weather, not climate. Here is how machine learning climate prediction works, what the evidence supports, and where it breaks.

Weather is not climate, and AI treats them differently

Weather is the state of the atmosphere over the next hours to roughly two weeks: whether it rains on Thursday, where a hurricane makes landfall. Climate is the statistics of weather over decades: how often Thursday rains, how hot the hottest day of the year gets, how those numbers shift as greenhouse gases accumulate. The two problems share equations but not goals.

A weather forecast is an initial-value problem. You start from the best estimate of the atmosphere right now and march it forward, and the answer is judged against what actually happens. Small errors in the starting state grow quickly, which is why forecasts lose skill after about two weeks. For a deeper look at why, see the butterfly effect and chaos.

A climate projection is a boundary-value problem. Nobody expects a climate model to say what the weather will be on March 3, 2071. Instead the model is run under assumed greenhouse gas levels, and scientists study the statistics of the output: averages, extremes, and how they change. The model has to stay stable and physically sensible for decades of simulated time, which is a much harder test for a neural network than a 10-day sprint.

Most headline results for AI belong to the first category. Weather models are trained and scored on short forecasts that can be checked against reality every day, so there is abundant evidence. Climate projections cannot be checked that way, because the future has not happened. That asymmetry shapes everything below.

The surprising number: GraphCast's authors report a 10-day global forecast in under one minute that outperforms the most accurate operational deterministic systems on 90% of 1,380 verification targets (Lam et al., arXiv:2212.12794).

What a physics-based climate model actually does

A traditional climate model, formally a general circulation model (GCM), divides the atmosphere and ocean into a three-dimensional grid and solves equations for fluid motion, heat, moisture, and radiation in each cell. Think of it as a very detailed accounting system: for every box, it tracks what flows in, what flows out, and what changes inside, then repeats for the next time step.

The grid is the catch. The Coupled Model Intercomparison Project (CMIP), which coordinates the simulations that feed international climate assessments, runs models whose atmospheric grid cells are typically on the order of 100 km wide. A recent AI emulator paper describes its reference model as having 100 km resolution, which is in that range. A thunderstorm, a sea breeze, or a mountain valley is far smaller than one cell, so the model cannot simulate it directly.

Modelers handle this with parameterizations: hand-built rules that estimate the average effect of processes too small to resolve, such as cloud formation or convection. Parameterizations are a major source of uncertainty in climate models, because they are approximations tuned against observations. Many of the AI methods discussed below aim at exactly this weak point.

Why they take so long to run

Each simulated time step, often around half an hour of model time, requires an enormous number of arithmetic operations. A century contains about 1.75 million half-hour steps, and that count is multiplied again by the number of scenarios and ensemble members scientists want to run. Exact wall-clock times vary widely by model, resolution, and computer, so a single figure such as "two weeks per run" would be misleading. The practical consequence is the one climate scientists describe: sweeping through many scenarios, or running many perturbed copies of the same scenario, is expensive enough to limit what questions get asked.

An AI emulator attacks the cost directly. Its authors, in one early example discussed below, report a roughly 100-fold reduction in wall-clock time and energy against the model it imitates. If that holds more broadly, it changes how many experiments a research group can afford.

How a neural network learns the atmosphere

Instead of coding the physics, you show a neural network what the atmosphere did, many thousands of times, and train it to predict the next state. The standard training ingredient is reanalysis data.

Reanalysis: the training set

Reanalysis blends a weather model with worldwide observations from satellites, balloons, aircraft, ships, and ground stations to produce a complete, consistent record of the atmosphere on a regular grid, with no gaps. ECMWF's ERA5 is the most widely used example. According to the Copernicus Climate Data Store, ERA5 provides hourly estimates from 1940 onward, with atmospheric fields at 0.25 by 0.25 degrees, and it is updated daily with about a five-day latency.

That is a spectacular training set: more than eight decades of hourly global snapshots. It also carries a quiet caveat. Reanalysis is not raw observation. It is a physics model nudged toward measurements, so every model trained on it inherits the biases of that physics model, as well as its strengths.

The grid and the rollout

A 0.25-degree global grid has 1,440 points around each latitude circle and 721 latitude lines, which is just over one million points. At each point the model tracks several variables (temperature, wind components, humidity, geopotential, and more) at many heights. The network takes the current state and, optionally, the previous one, and outputs the state a fixed time later.

To forecast further out, the model feeds its own output back in. GraphCast steps forward in six-hour increments, so a 10-day forecast means 40 consecutive predictions. This is called an autoregressive rollout, and it behaves like a game of telephone: a small error at step five gets passed along and amplified by step thirty. Training methods that penalize errors over multiple steps are what keep the rollout from drifting into nonsense.

The architecture matters as well. GraphCast represents the globe as a graph, with grid points as nodes linked to their neighbors, and passes messages along those links so information can travel across the planet. A flat image network would distort the poles, where longitude lines converge. Other models use transformers, the attention-based architecture behind modern language models (see transformers explained simply), applied to patches of atmospheric data. The general idea of learning layered patterns from data is covered in neural networks and deep learning.

The weather models: what each one claims

Several research groups released data-driven global weather models between 2022 and 2024. Their headline claims, taken from the authors' own abstracts, are summarized below. These are developer-reported benchmarks, not independent audits, and each paper uses a slightly different scoring setup.

Model	Year first posted	Resolution	Headline claim from the paper	Source
FourCastNet (Pathak, Anandkumar and colleagues)	2022	0.25°	Week-long forecast in under 2 seconds; matches IFS at short lead times on large-scale variables, better on fine-scale fields such as precipitation	arXiv:2202.11214
GraphCast (Lam et al.)	2022	0.25°	10-day forecast in under 1 minute; beats leading operational deterministic systems on 90% of 1,380 targets	arXiv:2212.12794
Pangu-Weather (Bi et al.)	2022	0.25°	First AI method to beat state-of-the-art numerical prediction on RMSE and ACC for all factors from 1 hour to 1 week, per its authors	arXiv:2211.02556
GenCast (probabilistic)	2023	0.25°	Beats ECMWF ENS on 97.4% of 1,320 targets; 15-day ensemble in about 8 minutes	arXiv:2312.15796
Aurora (Microsoft)	2024	multiple	Outperforms operational forecasts for air quality, ocean waves, and tropical cyclone tracks	arXiv:2405.13063

Two details in the table are worth slowing down for. First, GenCast is an ensemble model. A single forecast is one guess; an ensemble runs many slightly different versions to estimate the spread of possible outcomes, which is what allows a forecaster to say "70 percent chance" rather than "it will rain." The fact that a learned model can produce a calibrated ensemble was an open question until models like GenCast.

Second, the Aurora abstract reports training on over a million hours of diverse data and fine-tuning "at only modest computational cost." That is the foundation model approach: pretrain one large model on varied earth system data, then adapt it to specific tasks rather than building a new system each time. ClimaX, from Nguyen, Brandstetter, Kapoor, Gupta, and Grover, takes the same route. It was pretrained on climate datasets derived from CMIP6 and fine-tuned for forecasting and climate projection benchmarks (arXiv:2301.10343).

A note on how scoring works

Forecast skill is usually measured with root mean squared error (RMSE), which is the typical size of the gap between prediction and reality, and anomaly correlation coefficient (ACC), which measures whether a model gets the pattern of departures from normal right. A model can post a good RMSE by smoothing: predicting blurry, average-looking fields that are rarely badly wrong but also never capture the sharp feature that matters, such as a hurricane's intense core. This is why the tropical cyclone results in Aurora's paper, and the extreme-event tests discussed later, carry more weight than the global averages.

From research to operations: ECMWF's AIFS

The clearest sign that these models have left the lab is ECMWF's Artificial Intelligence Forecasting System (AIFS). According to ECMWF, the single-forecast version went operational at the end of February 2025, and the ensemble version, AIFS ENS, followed on July 1, 2025.

ECMWF's announcement reports that AIFS ENS produces 51 forecasts per run, outperforms state-of-the-art physics-based models on many measures, with gains of up to 20 percent for surface temperature, and runs at 31 km resolution compared with 9 km for the physics-based ensemble. It generates forecasts more than 10 times faster, and ECMWF puts the reduction in energy consumption at roughly 1,000 times (ECMWF, July 2025).

One sentence in that announcement deserves emphasis: the initial conditions still come from physics-based data assimilation. The AI system does not look at satellites and balloons directly. It takes the starting picture produced by the physics-based pipeline and runs it forward. In other words, the physics models did not become obsolete. They moved upstream, and the AI took over the most computationally expensive middle step.

Hybrid models: keeping the physics where it matters

If a purely learned model drifts over long runs, one remedy is to keep the parts of the physics that are well understood and learn only the parts that are not. This is the logic of NeuralGCM, from Kochkov and colleagues.

NeuralGCM pairs a differentiable numerical solver for large-scale atmospheric dynamics with machine learning components that replace the hand-tuned parameterizations. Differentiable here means the solver is written so that the training algorithm can push error signals back through it, which lets the learned pieces be trained while the physical solver stays in the loop. Per the authors' abstract, it is competitive with pure ML weather models on 1 to 10 day forecasts and with the ECMWF ensemble on 1 to 15 day forecasts, and it reports orders of magnitude of computational savings over conventional GCMs. With prescribed sea surface temperatures, it tracks global mean temperature over multiple decades, and its 140 km climate simulations show emergent features such as realistic frequency and trajectories of tropical cyclones (arXiv:2311.07222).

Why does this matter? The authors point out that earlier ML models matched GCMs on deterministic forecasts but had not shown improved ensembles or enough stability for long simulations. A hybrid addresses stability because the solver enforces the large-scale dynamics, much as guardrails keep a car on the road even when the driver wobbles. The cost is complexity: hybrid models are harder to build than a pure neural network and are still tied to the numerical solver's resolution.

Climate emulators and the projection problem

An emulator is a fast stand-in for a slow model. You run the expensive physics model to generate training data, then teach a network to reproduce its behavior. The analogy is a flight simulator trained from recordings of real flights instead of from aerodynamics equations: it can be fast and convincing within the conditions it has seen, and unreliable outside them.

The AI2 Climate Emulator (ACE) is a recent example. Its authors describe a roughly 200-million-parameter autoregressive model that emulates an existing 100 km global atmospheric model, stays stable over 100 years of simulation, nearly conserves column moisture without explicit constraints, and beats a strong baseline on over 90 percent of tracked variables. They claim nearly 100 times less wall-clock time and roughly 100 times better energy efficiency than the reference model on typical resources, and that it handles a previously unseen historical sea surface temperature dataset without fine-tuning (arXiv:2310.02074).

The key limitation is built into the word emulator. ACE reproduces a physics model, not the real atmosphere, so its ceiling is the quality of that model. If the reference model has a cloud bias, the emulator copies it, quickly. What emulators offer is speed for sensitivity testing: run a thousand slightly different scenarios in the time it used to take to run a few.

Can AI project climate to 2100?

Not on its own, and no one has shown otherwise. A projection to 2100 means simulating a climate with greenhouse gas concentrations the training data has never contained. A network that learned the relationship between ocean temperatures and rainfall from 1940 to 2023 has no guarantee that the relationship survives in a world several degrees warmer. Physics, by contrast, is designed to extrapolate: conservation of energy does not depend on the century. This is why hybrid and emulator approaches are being developed as accelerators for physics-based projection rather than replacements.

Downscaling: from 25 km to 2 km

Global models are coarse, and cities are small. Downscaling translates coarse output into local detail, and it is one of the places where machine learning has a natural fit, because the task resembles image super-resolution.

A notable example is CorrDiff, a residual corrective diffusion model. In its paper, a UNet predicts the mean and a diffusion model predicts the residual detail, producing 2 km output over Taiwan from 25 km global reanalysis inputs. The authors frame it as a cost-effective machine learning alternative to expensive kilometer-scale simulations (arXiv:2309.15214). Diffusion models are the same family used by AI image generators, covered in how diffusion models work.

There is a trap here that is easy to miss. A downscaled map looks sharp, but sharpness is not the same as information. The finer detail is learned from training examples, so it reflects the patterns of the region and period the model saw. A neural network can render a believable street-level rainfall pattern that the coarse input does not actually determine. For decisions such as flood planning, the right question is how the downscaled output was validated against independent local observations, not how crisp it looks.

Where machine learning climate prediction fails

The failure modes below are the reason experts treat AI as an accelerator and not as a replacement.

Extremes outside the training record

One direct test comes from a 2025 study in the Proceedings of the National Academy of Sciences. The researchers trained versions of the FourCastNet model on ERA5 data, with one version excluding the strongest tropical cyclones, and then tested them on Category 5 storms, the so-called gray swans: extreme events that were not in the training data but are physically possible.

The version trained without strong storms could not accurately forecast Category 5 cyclones, even though all versions scored similarly on ordinary global weather. In other words, the model did not extrapolate from weaker storms to stronger ones. The authors argue that the finding probably extends to other AI weather and climate models because they use similar learning strategies, and they call for new learning approaches before such models are trusted for early warning of unprecedented events (Sun et al., PNAS, 2025, DOI 10.1073/pnas.2420914122).

A warming climate produces more events of exactly this kind. This is the central reason why the question "can AI forecast climate?" splits into a yes for the weather we have seen and a not-yet for the weather we have not.

Inherited bias and the reanalysis ceiling

Because reanalysis is physics-model output nudged by observations, models trained on it inherit its systematic errors. An emulator trained on a physics model's output inherits that model's errors even more directly. Learning from data does not remove bias; it can freeze it in place and run it faster.

Physical consistency

A neural network does not automatically conserve mass, energy, or moisture. It can drift slowly, and over a century-long run, small violations compound. ACE's authors note that their model nearly conserves column moisture without explicit constraints, which is notable precisely because it is not guaranteed. Physics-informed approaches, which build conservation into the architecture or the loss function, are an active area; see physics-informed neural networks.

Interpretability

When a physics model produces a surprising result, scientists can trace which process caused it. A neural network offers fewer such handles. Explainability methods exist, but they describe which inputs the model weighted, not why the atmosphere behaves that way. For a field whose purpose is understanding as much as prediction, that is a real cost.

Chaos does not go away

No model, learned or physical, can forecast the specific weather of a day three weeks out, because small initial uncertainties grow. AI can produce a forecast faster and may reduce certain errors, but the underlying predictability limit comes from the atmosphere. Turbulent fluid behavior is the reason; see why turbulence is hard to predict and neural networks and fluid dynamics.

Summary table: where each approach stands

Task	Physics-based models	Pure ML models	Hybrid / emulator
Medium-range weather (1 to 15 days)	Operational standard, slower	Matches or beats on many measures, per developer papers and ECMWF	NeuralGCM competitive with ECMWF ensemble
Probabilistic forecasts	51-member ensembles at 9 km (ECMWF)	GenCast and AIFS ENS produce ensembles	Under development
Extreme events never seen in training	Physics generalizes in principle, but resolution limits intensity	Failed in the 2025 gray swan test	Untested at scale
Multi-decade climate simulation	Standard method	Not established	ACE stable for 100 years as an emulator; NeuralGCM tracks decades with prescribed SSTs
Local-scale downscaling	Costly regional models	Diffusion downscaling shows promise	Validation against local observations is essential

What this means

The strongest evidence for machine learning in this field sits in short-range forecasting, where models can be graded daily against the real atmosphere and the operational centers have been convinced enough to run them. The weaker evidence, and the more important questions, sit in climate projection, where the test is the future and the training data stops at the present.

The sensible reading is a division of labor. Physics supplies the structure, the initial conditions, and the ability to extrapolate beyond experience. Machine learning supplies speed, which turns a handful of expensive simulations into thousands of cheap ones, and a new way to correct the approximations that physics models make at small scales. The open scientific question is how far that speed can be trusted when the climate leaves the range of anything it has learned from. For the wider picture of how warming works, start with climate change science; for AI applied to hazards, see AI and disaster risk assessment.

FAQ

Q: Will AI replace traditional climate models? A: Current evidence points to a partnership. ECMWF's operational AI system still takes its starting conditions from physics-based data assimilation, and hybrids such as NeuralGCM keep a physical solver in the loop. Physics models remain the way to explain why the climate changes and to extrapolate to conditions with no historical analog.

Q: How much data does an AI weather or climate model need? A: The leading models learn from reanalysis, which for ERA5 means hourly global fields from 1940 onward. Gathering the data is not the barrier, since it is free to download. The cost is in training: large models require substantial GPU or TPU resources, though exact figures differ by model and are rarely published in full.

Q: Can AI improve climate projections to 2100? A: Not by itself. A projection to 2100 requires simulating greenhouse gas levels outside the training record, and the 2025 gray swan study shows that standard models struggle with events they have not seen. Emulators such as ACE can make existing physics-based simulations much cheaper to run, which indirectly improves how thoroughly uncertainty can be explored.

Q: Is there a risk of overfitting to recent climate trends? A: Yes. A model trained on the satellite era learns the climate of that period, and relationships may shift as the climate warms. Hybrid designs, tests on held-out extremes, and comparison against physics-based runs are the main safeguards researchers use.

Q: How do AI models handle uncertainty? A: Through ensembles. GenCast and ECMWF's AIFS ENS generate many forecasts from slightly different starting points or noise samples and report the spread. AIFS ENS produces 51 forecasts per run. Single deterministic AI forecasts, by contrast, give no direct measure of confidence.

Q: What is the computational cost of running these models? A: Once trained, a forecast is cheap. GraphCast's authors report under a minute for a 10-day global forecast, FourCastNet's authors report under two seconds for a week, and ECMWF reports roughly 1,000 times lower energy use for its ensemble. Training is the expensive part and is not captured in those figures.

Q: Can AI predict climate tipping points? A: There is no demonstrated method. Tipping points, such as the collapse of major ocean circulation patterns, have few or no historical examples in the instrumental record, which is exactly the situation where data-driven models are weakest. Research on early-warning signals is ongoing, but it should be treated as exploratory.

Sources

• Lam et al., GraphCast: arxiv.org/abs/2212.12794
• Pathak et al., FourCastNet: arxiv.org/abs/2202.11214
• Bi et al., Pangu-Weather: arxiv.org/abs/2211.02556
• GenCast: arxiv.org/abs/2312.15796
• Kochkov et al., NeuralGCM: arxiv.org/abs/2311.07222
• ACE climate emulator: arxiv.org/abs/2310.02074
• Aurora: arxiv.org/abs/2405.13063
• ClimaX: arxiv.org/abs/2301.10343
• CorrDiff downscaling: arxiv.org/abs/2309.15214
• ECMWF, AIFS ENS operational: ecmwf.int
• Copernicus Climate Data Store, ERA5: cds.climate.copernicus.eu
• Eyring et al., CMIP6 design: gmd.copernicus.org
• Sun et al., "Can AI weather models predict out-of-distribution gray swan tropical cyclones?", PNAS, 2025, DOI 10.1073/pnas.2420914122

---

Read next:

• AI-Powered Carbon Capture Technology — Can Machine Learning Scale Carbon Removal?
• Smart Cities and AI — How Machine Learning Reduces Urban Energy Waste
• How AI Optimizes Renewable Energy Grids — Machine Learning for Smart Grids