Modeling extreme weather events, the kind that cause the most damage, is computationally brutal. A 1-in-100-year heatwave requires at least 100 simulated years to generate even a single example by brute force. With high-resolution climate models, the computing time and energy cost associated with such a simulation would be measured in months or years, even on supercomputers.
Researchers at the University of Chicago, CNRS Paris, and New York University have developed a method that cuts this cost dramatically. Their AI-boosted rare-event sampling (AI+RES) framework, accepted for publication in Physical Review Letters, combines a deep-learning weather emulator with a trajectory-splitting algorithm to achieve up to a 1,000-fold reduction in computational resources for characterizing extreme heatwave statistics.
How It Works
The framework has two components. The first is an AI weather emulator, a deep neural network trained on climate model output that can run ensemble forecasts at near-zero computational cost. The emulator functions as a “score function,” predicting which simulation trajectories are most likely to lead to an extreme event.
The second component is a trajectory-splitting rare-event sampling method. At scheduled resampling times, the algorithm duplicates promising trajectories (those the AI identifies as likely to produce an extreme event) and terminates unpromising ones. Only the most promising trajectories are then passed to the full physics-based climate model, in this case, PlaSim, for high-fidelity simulation.
The AI component solves a longstanding problem with rare-event sampling: designing a good score function traditionally requires deep domain expertise and extensive trial and error, and it is especially difficult for short-term extremes like heatwaves. The AI learns the score function automatically from the climate model output.
Demonstrated Performance
The team tested AI+RES on mid-latitude heatwaves in two regions: centered over France and the U.S. Midwest. The framework reproduced the ground-truth statistics from long PlaSim simulations at dramatically lower cost.
Standard RES without the AI booster failed entirely for the rarest events, it could not produce a single example of the most extreme heatwaves. Pure AI models without the physics component were inaccurate and could not extrapolate beyond their training data, a limitation of purely data-driven weather prediction.
The Physics World article reporting the results says the approach achieved “up to 1,000-fold” computational savings. The paper itself reports 30 to 300 times lower cost for the specific PlaSim heatwave validation, with the higher figure reflecting the AI+RES combination. The discrepancy reflects the difference between a round accessible figure for a general audience and the specific measured range.
The method yields both accurate statistics and physical insights into the mechanisms driving extreme events, meaning it can be used not just to predict heatwave frequency but to understand why they occur.
Why It Matters
Climate models are becoming more detailed and more expensive to run. The computational cost of high-resolution models limits how many simulations researchers can perform, which in turn limits their ability to estimate the probability of extreme events that may be rare but devastating.
If the AI+RES approach can be generalized to other types of extreme events, tropical cyclones, atmospheric rivers, floods, severe thunderstorms, it could fundamentally change how climate risk is assessed. Rather than relying on statistical extrapolation from limited data, models could directly simulate thousands of years of extreme events at a fraction of the current cost.
The authors, co-first authors Amaury Lancelin (CNRS/ENS Paris) and Alexander Wikner (University of Chicago), with corresponding authors Dorian Abbot (Chicago), Freddy Bouchet (ENS Paris), Pedram Hassanzadeh (Chicago), and Jonathan Weare (NYU), have made the AI+RES code available for other researchers to adapt to their own climate models.
Caveat
The method has been validated on only one climate model (PlaSim) and one class of events (mid-latitude summer heatwaves). The 1,000-fold figure is an aspirational upper bound more reflective of the Physics World coverage than the measured range in the paper itself, which reports 30-300 times savings for the heatwave case. Generalization to other event types and higher-resolution models is the next step.
Disclosure: Based on a paper accepted in Physical Review Letters. arXiv: 2510.27066. DOI: 10.1103/b1gc-9c2q. Reported via Physics World, July 6, 2026.