Does Aurora Encode Atmospheric Structure? Latent Regime Analysis and Attribution

2026-06-24

AI Summary

This study investigates whether the Aurora weather foundation model spontaneously encodes atmospheric structural information without explicit supervision, with a focus on its ability to distinguish between seasonal cycles and extreme storm events. Through spatial pooling PCA, layer-wise relevance propagation (LRP), and region-specific masking perturbation experiments, the authors analyze the model’s latent representational structure. The work reveals for the first time that Aurora implicitly learns atmospheric vertical structure and meteorological coherence, with its latent space predominantly governed by seasonal periodicity. Although extreme storms do not form linearly separable clusters, LRP successfully identifies key features aligned with the three-dimensional structure of the 1987 Great Storm; masking this region degrades forecast performance by a factor of 3.31, underscoring the model’s reliance on physically consistent representations.

Abstract

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.

Problem

• atmospheric structure
• latent representation
• foundation model
• extreme weather events
• seasonal cycles

Innovation

• latent space analysis
• layer-wise relevance propagation
• atmospheric dynamics
• foundation models
• meteorological coherence