AI and the World's Hard Problems · Earth & climate science · arXiv · 2026-10-19
Rescene: stable century-long climate emulation from a frozen weather model
Machine-learning weather models drift, blow up, or lose variability when run freely for years, and dedicated climate emulators are costly to train. The authors wrap a frozen weather model with a small 0.4M-parameter component called Rescene, which anchors it to a climatology and adds a band-limited stochastic perturbation. A 100-year, 8-member run showed no detectable drift, and the 16-member ensemble on the 2022 test year was calibrated. The free run also reproduced 71% of the observed blocking frequency, with all constants fitted on a single training year.