Tropical Cyclons: Parsimonious autoencoders show how GraphCast represents position, intensity and pressure

By surveying the internal states of GraphCast on 205 cyclones, a team shows that each physical attribute is distributed over several characteristics, according to an organization that repeats itself from layer to layer.

AI-based weather models are now good at predicting extreme events, but little is still known about how they represent them internally. Jiapei Liu, Ruidong Li, Ting Sun, and Guangheng Ni are tackling this question with sparse autoencoders (SAEs), an interpretability technique that breaks down hidden states in a network into more readable features. Their case study: tropical cyclones in GraphCast, Google's DeepMind model. They analyzed 2,115 six-hour time steps covering 205 cyclones, comparing activations with the systems' position, intensity, and pressure structure, on five decompositions of the eighth layer and on the 16 layers of the processor.

The result: no attribute is carried by a single characteristic, each one is distributed over several of them, and the position and the pressure structure share a smaller subset. This organization is found from one independently trained autoencoder to another and from one layer to another, without the need to align the characteristics. More intriguingly, in the deep layers, the main characteristic related to the position is better at the position and the pressure structure six hours later than at the present moment. For the authors, this organization offers a reproducible level of description to interpret what GraphCast "knows" about cyclones.