Tropical Cyclone Attributes in Sparse Autoencoder Features of an AI Weather Model
Abstract
Artificial intelligence (AI) weather models now forecast weather events skillfully, but how they represent these events internally remains unknown. Sparse autoencoders (SAEs) provide sparse features for probing hidden states, but how information about multiple physical attributes of an event is organized across them remains unclear. Using tropical cyclones in GraphCast as a test case, we analyze 2,115 six-hourly timesteps from 205 cyclones, relating SAE activations to position, intensity, and pressure structure across five Layer-8 decompositions and 16 processor layers. Each attribute relates to multiple features; position and pressure-structure relations overlap in a smaller subset. This organization recurs across independent SAEs and layers without feature alignment. In deeper layers, the leading position feature aligns better with position and pressure structure six hours ahead than at present. These results suggest that the organization of physical-attribute relations provides a reproducible level of description for interpreting cyclone-related information in GraphCast.
Related
- Skillful Seasonal Tropical Cyclone Prediction with AI Weather--Climate Models Forced by Dynamically Predicted SSTs
- Butterfly Effect Confirmed in Global AI Weather Models: Evidence from Tropical Cyclone Forecasting
- Enhancing Tropical Cyclone Prediction Through AI Weather Model Ensemble Forecasts and Multi‐Source Data Integration
- A Hybrid Deep Learning Framework for Tropical Cyclone Intensity Estimation With Spatiotemporal Features and Physical‐Persistence Adjustment
- Tibetan Plateau uplift increased western North Pacific tropical cyclone frequency
BibTeX
@misc{liu2026tropical,
title = {{Tropical Cyclone Attributes in Sparse Autoencoder Features of an AI Weather Model}},
author = {Jiapei Liu and Ruidong Li and Ting Sun and Guangheng Ni},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.28904},
eprint = {2610.28904},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.28904},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Tropical Cyclone Attributes in Sparse Autoencoder Features of an AI Weather Model
AU - Jiapei Liu
AU - Ruidong Li
AU - Ting Sun
AU - Guangheng Ni
PY - 2026
DA - 2026/10/20
PB - arXiv
DO - 10.48550/arxiv.2610.28904
UR - https://arxiv.org/abs/2610.28904
AB - Artificial intelligence (AI) weather models now forecast weather events skillfully, but how they represent these events internally remains unknown. Sparse autoencoders (SAEs) provide sparse features for probing hidden states, but how information about multiple physical attributes of an event is organized across them remains unclear. Using tropical cyclones in GraphCast as a test case, we analyze 2,115 six-hourly timesteps from 205 cyclones, relating SAE activations to position, intensity, and pressure structure across five Layer-8 decompositions and 16 processor layers. Each attribute relates to multiple features; position and pressure-structure relations overlap in a smaller subset. This organization recurs across independent SAEs and layers without feature alignment. In deeper layers, the leading position feature aligns better with position and pressure structure six hours ahead than at present. These results suggest that the organization of physical-attribute relations provides a reproducible level of description for interpreting cyclone-related information in GraphCast.
ER -