Watch an AI Weather Model Learn (and Unlearn) Tropical Cyclones

Rebecca Baiman, Ankur Mahesh, Elizabeth A. Barnes

arXiv preprint · 20 March 2026 · Colorado State University (incl. CIRA), Lawrence Berkeley National Laboratory / NERSC, University of California, Berkeley

Paper PDF

In one sentence. Checkpoint-by-checkpoint analysis shows a Spherical Fourier Neural Operator learns tropical cyclone intensity for some storms, then unlearns it, linked to moist environments.

Abstract

In a changing climate, artificial intelligence (AI) weather models have the potential to provide cheaper, faster, and more accurate forecasts of high-impact weather events. To realize this potential and gauge trustworthiness, there is a need for more research on how models learn extreme events and how that learning might be improved. Here, we investigate how a Spherical Fourier Neural Operator (SFNO) learns tropical cyclones (TCs) by saving every checkpoint from training and analyzing storm specific metrics. We find evidence that for some storms the SFNO learns information about TC intensity that it loses later in training. This unlearning pattern is associated with anomalously moist environments and may be due to the model unlearning the relationship between moisture and TC intensity. This work provides a first example of leveraging task-specific training dynamics to further our understanding of how AI weather models learn extreme events.

Labels

Labels and summary are generated by Claude and reviewed weekly. Wrong? Edit this entry.