Does Aurora Encode Atmospheric Structure? Latent Regime Analysis and Attribution
For researchers and practitioners using ML weather models, this work provides interpretability insights into a black-box foundation model, but the findings are incremental as they apply existing analysis methods to a single model.
The paper investigates whether the Aurora foundation model encodes atmospheric structure, finding that its latent space is organized by seasonal cycles and that the model attends to 3D vertical storm features, with perturbation tests showing masking relevant regions degrades forecasts 3.31× more than random masking.
ML foundation models are able to emulate atmospheric dynamics accurately and efficiently but operate as opaque ``black boxes''. We investigate the internal representations of the Aurora model using spatially pooled PCA and layer-wise relevance propagation (LRP). We find evidence that Aurora's latent space is primarily organized by seasonal cycles, whereas extreme storm events do not form a linearly separable cluster. LRP indicates that the model attends to features consistent with the 3D vertical structure of the Great Storm of 1987. Perturbation tests show masking relevant regions degrades forecasts $3.31\times$ more than random masking. These findings suggest that Aurora learns meteorological coherence and vertical structure without explicit instruction.