AI helps physics models predict rare extreme heatwaves

What happened?

A team from the Laboratoire de meteorologie dynamique, the University of Chicago, New York University and RTE has published a new way to study rare extreme weather events such as severe heatwaves. The method is called AI+RES: artificial intelligence plus rare-event sampling.

The problem is that the most dangerous weather events are rare. A very severe heatwave might have a return period of hundreds or thousands of years in a particular model climate. Waiting for enough examples in real observations is impossible, and running a detailed physics-based climate model for many thousands of simulated years can be too expensive.

The researchers used PlaSim, an intermediate-complexity global climate model, as the physics-based model for their proof of concept. They trained an AI weather emulator on model data, then used it to guide which simulated atmospheric paths were more likely to lead towards an extreme heatwave.

The result was not simply an AI forecast replacing physics. The AI acted as a fast guide, while the physics model still generated the final heatwave examples and statistics. In tests over France and the US Midwest, AI+RES estimated very long return-period heatwaves much more efficiently than direct simulation with the same computing budget.

The simple version

Think of a climate model as a very detailed physics simulation of the atmosphere. It repeatedly applies equations for motion, energy transfer, radiation, pressure, moisture and temperature on a grid covering the Earth.

If you want to study an ordinary warm day, the model will produce plenty of examples. If you want to study an extreme heatwave that is very unlikely, most model runs will be unhelpful because they never reach the event you care about.

Rare-event sampling is a smarter search strategy. Instead of treating every simulated path equally, it makes more copies of paths that seem to be heading towards the rare event and drops paths that seem less useful. This still has to be done carefully so the final probability estimate is not biased.

The new idea is to use AI to choose the score function: the clue that tells the rare-event algorithm which paths look promising. The AI forecast is cheap and fast, so it can look ahead many times. The physics model then follows the selected paths to produce realistic heatwave cases.

Worked equations

Why it matters

Extreme heat is a physics problem as well as a social problem. Temperature, humidity, radiation, air motion and land-surface conditions all affect how severe a heatwave becomes and how dangerous it is for people, crops, transport and electricity systems.

For planners, the rarest events are often the most important. A power grid might cope with ordinary hot days but struggle with a once-in-many-decades heatwave because electricity demand rises, cooling systems work harder and power lines can sag as they expand.

The method produced long-return-period estimates with far less direct simulation than brute force modelling. The authors report that AI+RES could study very rare events at around two orders of magnitude lower computational cost, with speedups above 100 in some tested return-period ranges.

This matters because scientists need both probabilities and physical examples. A probability tells us how often a kind of event may occur in the model. A full simulated trajectory shows how pressure patterns, soil moisture and atmospheric motion built up before the heatwave.

Physics you already know

The first school link is thermal physics. OCR Module 5 and AQA Paper 2 both include temperature, internal energy, ideal gases and energy transfer. Heatwaves are a real-world situation where a small change in average temperature can represent a very large energy change across the atmosphere and surface.

The atmosphere is not a simple sealed ideal gas, but the exam idea still helps: temperature is linked to the average kinetic energy of particles. When air and surfaces warm, energy has been transferred into microscopic motion and interactions.

This is also a strong example of data analysis. The researchers did not just ask whether one model run made a hot week. They compared distributions, return-period curves, uncertainty bands and maps of the atmospheric state before and during the heatwave.

The modelling is an example of using a simplified system wisely. PlaSim is less detailed than the most expensive climate models, but it is useful for testing a method because researchers can generate a large reference dataset and check whether the new algorithm reproduces the correct statistics.

Pupils sometimes hear AI described as if it replaces science. Here it does something more interesting: it helps choose where the physics model should spend its effort. The physical model is still needed to keep the heatwave examples tied to equations of motion and energy transfer.

Science ideas to understand

A Level stretch

The algorithm uses rare-event sampling, a form of importance sampling. In ordinary direct sampling, most simulations miss the tail of the distribution. In rare-event sampling, the algorithm deliberately explores the tail while keeping weights so the probability estimate remains unbiased.

The AI emulator is not trusted blindly. In the paper, AI-only direct sampling produced biased estimates for rare heatwaves. The useful role of the AI was ranking which current atmospheric states were more likely to lead to the target event, not replacing the final physics simulation.

The study used a stationary climate setup for validation. That means the researchers were testing the method under controlled conditions, not making a final real-world forecast for future climate change in a particular country.

The authors also note that land-surface processes such as soil moisture can strongly affect surface temperature extremes. Dry soil can reduce evaporative cooling, making air temperatures climb more easily during blocking weather patterns.

A future challenge is applying AI+RES to more detailed climate models and more complex compound events, such as heat plus drought or high electricity demand plus low renewable generation.

Key words

Quick pupil questions