Minisymposium Presentation

Foundation models have achieved remarkable accuracy in short- to medium-range weather forecasts, primarily focusing on atmospheric variables. However, predicting new physical variables typically requires training or fine-tuning the model with additional datasets, incurring significant costs. We show that new variables can be learned directly from the latent space of the Aurora foundation model. Our frugal extension involves training lightweight decoders using a small dataset, specifically a subset of ERA5 and MSWEP. These decoders accurately predict surface variables related to the water cycle, establishing the first baseline for many of these variables. For precipitation, our decoder achieves results comparable to those in the literature. This work presents an affordable method to extend foundation models beyond atmospheric predictions. It also suggests that Aurora captures an internal representation of the Earth system, contributing to a better definition and understanding of foundation models.