AI+Physics hybrid framework: Optimizing climate simulation compute with AI predictions

bravo_abad · x · 2026-08-07

Amaury Lancelin and coauthors introduce AI + RES, a hybrid framework that uses fast AI weather forecasts to decide where expensive physics-based simulations should spend their computational budget.

The approach is pragmatic: instead of asking the AI emulator to reproduce the statistics of extremely rare events directly, it is used as a ranking function inside a rare-event sampling algorithm. At intermediate times, the AI predicts which simulated trajectories are most likely to evolve into extreme heat waves. Promising trajectories are duplicated and unpromising ones discarded, while the underlying events are still generated by the physics-based climate model. This avoids the notorious unreliability of AI extrapolation for rare events.

Original post →

More from Research

Research channel →