AI and the World's Hard Problems · Earth & climate science · arXiv · 2026-10-20
Interpretability of AI weather models via intermediate decoding
AI weather models are hard to interpret because their internal representations are not physically meaningful. The authors decode intermediate latent states back into meteorological fields, using affine transformations to align them with the decoder's input space. Applied to GraphCast, Aurora and ArchesWeather, the method shows forecast skill improving gradually through the processor layers. The models also refine forecasts from coarse to fine spatial scales, and the encoder-decoder plays a significant role in forecast construction.