Open the black box of GraphCast, Aurora and ArchesWeather: decoding the intermediate layers shows a forecast that refines from coarse to fine

Meteorologists in Munich re-use the model decoder to read their internal states during calculation: quality progresses layer after layer, from large scales to small ones.

AI weather forecasting models often beat physical models, but we don't know how they build their forecasts: their internal representations, which are very large, do not make direct meteorological sense. Matthias Beylich, Kirsten Tempest, Tobias Selz and George Craig propose "intermediate decoding": taking the latent state at a given depth and passing it through the final decoder to translate it back into weather fields (pressure, wind, temperature, etc.). As each layer does not use the same base, they learn simple affine transformations to align these states with what the decoder expects.

Three models, one scheme

Applied to GraphCast, Aurora, and ArchesWeather, the method yields convergent results. The quality of the prediction progresses gradually through the processing blocks. Shorting the entire core of the network shows that the encoder-decoder pair doesn't just copy the input, it participates in the prediction itself. And looking at what each block adds reveals that all three models start with the large structures before refining the smaller-scale details.

Finally, the authors verify the physical consistency of these intermediate predictions: the large scales remain balanced from the start, while the small ones become a little more balanced with each block — a way to understand, and perhaps correct, what these models actually do.