This research evaluates the performance of the Aurora weather foundation model by using lightweight decoders to predict hydrological and energy variables not included in its original training. The study highlights that this decoder-based approach significantly reduces training time and memory requirements compared to fine-tuning the entire model, while still achieving strong accuracy. A key finding is that decoder accuracy is influenced by the physical correlation between the new variables and those initially used for pretraining, suggesting that Aurora's latent space effectively captures meaningful physical relationships. The authors argue that the ability to extend foundation models to new variables without full fine-tuning is an important quality metric for Earth sciences, promoting accessibility for communities with limited computational resources. They conclude that rich latent space representations allow for accurate predictions of new variables using lightweight extensions, advocating for future foundation models that encompass a broad range of physical processes.

Reference:

Lehmann, F., Ozdemir, F., Soja, B., Hoefler, T., Mishra, S., & Schemm, S. (2025). Finetuning a Weather Foundation Model with Lightweight Decoders for Unseen Physical Processes. arXiv preprint arXiv:2506.19088.