AI in Climate Change Prediction: Where It Wins and Fails

In the space of three years, AI in climate change prediction has gone from a research curiosity to a system that one of the world's leading forecasting centres runs every six hours. The papers behind that shift report striking wins, and they are real. What the headlines rarely put next to them is the second set of results: the places where the same models fall short, which happen to be the events a warming climate produces more of.

This piece puts the two sets side by side. It is written for the people who have to decide what to trust and what to fund: researchers, agency staff, climate investors and founders building on top of these models.

What AI in climate change prediction actually covers

Two different problems travel under the same phrase. Weather prediction asks what the atmosphere will do over the next days to two weeks. Climate prediction asks how its statistics will behave over decades, under forcing the past never saw. Both have traditionally been handled by general circulation models, which the authors of the NeuralGCM paper describe as physics-based simulators that combine a numerical solver for large-scale dynamics with tuned representations of small-scale processes such as cloud formation.

The AI models of the past three years learned mostly from reanalysis, a reconstruction of past weather. That origin explains both their strengths and their failures.

The wins: four models and what they measured

GraphCast: ten days in under a minute

GraphCast predicts hundreds of weather variables for the next ten days at 0.25 degree resolution in under a minute. In its Science paper it outperformed the most accurate operational deterministic systems on 90% of 1,380 verification targets, with better tracking of tropical cyclones, atmospheric rivers and extreme temperatures.

GenCast: the ensemble that beat ECMWF

A single forecast says nothing about uncertainty, and uncertainty is what warnings and energy planning depend on. GenCast generates an ensemble of 15-day global forecasts for more than 80 variables in 8 minutes, and in Nature it showed greater skill than ENS, the ECMWF ensemble widely treated as the world's best, on 97.2% of 1,320 targets.

NeuralGCM: a hybrid that runs for decades

NeuralGCM keeps a differentiable physics solver for atmospheric dynamics and hands the small-scale processes to machine learning. It is competitive with the ECMWF ensemble for one to fifteen days and, with prescribed sea surface temperature, tracks climate metrics for multiple decades. At 140 kilometre resolution it produces realistic tropical cyclone frequency and trajectories, at orders of magnitude lower computational cost than conventional models.

AIFS: the first operational open model

On 25 February 2025 ECMWF took the AIFS into operations, running side by side with its physics-based IFS. It gains up to 20% on tropical cyclone tracks and uses roughly 1,000 times less energy per forecast. Its first weeks in service included the Gulf Coast snow of 21 January 2025, an event with a return period of roughly 100 years that barely appears in the training data, which the AIFS Single captured. ECMWF read that as a sign the model had learned something about atmospheric physics rather than memorising its archive.

Four independent results, one direction. On average conditions, AI now matches or beats the best physics at a fraction of the cost.

The failures: records, storms and the monsoon

Record-breaking extremes

A 2026 study in Science Advances tested GraphCast, Pangu-Weather, Fuxi and two operational variants against ECMWF's physics-based HRES on record-breaking heat, cold and wind. HRES won consistently, across nearly all lead times, and the gap was largest at short range. The AI models underestimated both the frequency and intensity of records, and the further a record was exceeded, the larger the error. The authors describe the behaviour as an implicit soft cap near the most extreme value in the training period, 1979 to 2017. Physics-based models have no such bound, because the equations allow them to extrapolate.

This matters more than any single score. Record-breaking events are the ones early warning systems exist for, and the study notes they are particularly frequent in a rapidly warming climate. Its conclusion is unambiguous: AI models are not yet ready to be relied on alone for early warning and disaster management.

Storm peaks get smoothed

ECMWF's own forecasters reached a compatible view by a different route. Reviewing Storms Amy and Éowyn, they found the AIFS captured tracks, pressure patterns and synoptic structure well, often better than the IFS, but underestimated peak winds and smoothed the sharp fronts where the damage happens. The cause is partly design: a model trained to minimise mean squared error favours smooth, averaged fields. The same newsletter that praised the Gulf Coast snow forecast concedes that the AIFS Single still struggles with small-scale extremes.

Station data is harsher than reanalysis

Most headline scores are computed against reanalysis, the same kind of data the models were trained on. A Stanford team evaluated seven AI models, including GraphCast, AIFS and GenCast, against observations during the South Asian monsoon. The models handled large-scale dynamics with reasonable accuracy but showed substantially higher errors against ground stations than against reanalysis, and disagreed on extreme precipitation and monsoon cyclone tracks. For a farmer or a flood agency, the station is what counts.

Can a weather model predict a warmer climate?

This is the question the phrase "climate change prediction" ultimately asks, and the honest answer is: partly, and not yet reliably.

One encouraging test ran weather models trained on present-day data in pre-industrial and 2.9 degree warmer climates. They stayed skilful, which suggests the dynamics of daily weather do not change fundamentally as the planet warms. But two of the three models showed a global cold bias in the warmer state, drifting back toward the climate they were trained on.

Longer runs expose a second problem. Purely data-driven models tend to blur towards climatology within weeks or diverge until they blow up numerically. Hybrids like NeuralGCM are the exception so far, stable over decades, but its authors state plainly that it does not extrapolate to substantially different future climates. A century-scale projection is exactly that kind of extrapolation.

The division of labour

Put the wins and failures together and a working arrangement appears, one the most careful institutions have already adopted.

• AI for speed, scale and ensembles. Cheap forecasts mean far larger ensembles, better uncertainty estimates and forecasts in places that could never afford a supercomputer.
• Physics for records and the unprecedented. Where a forecast must go beyond anything in the training data, the equations still carry the load.
• Hybrids as the bridge. NeuralGCM and ECMWF's research into blending learned and physical components are where climate-scale prediction is most likely to improve.
• Verification against observations, not just reanalysis. The monsoon evaluation shows how much a score can flatter a model.

ECMWF describes the AIFS and IFS as complementary. The Science Advances authors go further and argue it remains vital to fund and run both in parallel. The applied case for that pairing already exists: blended AI forecasts reached 38 million farmers in India ahead of an unusual monsoon, a story we examine in AI in climate change adaptation. The emissions side of the ledger belongs to a separate piece on AI in climate change mitigation, and the wider map is in how AI can help in climate change.

Where the summit comes in

"Fund both in parallel" is easy to write and hard to arrange. Physics-based modelling depends on public institutions and supercomputing budgets; the AI models increasingly come from private laboratories. The hybrids that look most promising sit between the two, and so does their funding.

That gap is what the EX Future Summit is built around. It runs 18 to 20 November 2026 online, in Las Palmas and in Bali, as a single thirty-hour broadcast between two islands twelve hours apart, so a modelling group in Europe and a forecasting agency in Asia can sit in the same conversation. Its program includes a Global Financial Raise track and hosted matchmaking between universities and private partners: working sessions rather than keynotes, for exactly the kind of project that needs a research team and a capital partner at the same table.

FAQ

Is AI better than traditional climate models?

For medium-range weather on average scores, often yes: GraphCast, GenCast and the AIFS all beat leading physics systems on most targets. For record-breaking extremes and for climates very different from today's, physics-based models remain more reliable.

Which AI weather model is used operationally?

ECMWF's AIFS has been operational since 25 February 2025. It runs side by side with the physics-based IFS rather than replacing it, and ECMWF describes the two as complementary.

Why do AI models underpredict extreme heat?

They are trained on historical weather and largely interpolate between patterns they have seen. For record-breaking heat they behave as if capped near the most extreme training value, so the bigger the record, the bigger the underestimate.

Can AI run century-long climate projections?

Not reliably yet. Hybrid models such as NeuralGCM run stably for decades when sea surface temperatures are prescribed, but their authors note they do not extrapolate to substantially different future climates, and purely data-driven models tend to drift or become unstable over long runs.