Abstract

Reliable understanding of subseasonal monsoon variability is important for climate-risk management, agriculture, water resources, and disaster preparedness. While foundation models are increasingly used for weather and climate applications, it is still unclear what types of large-scale climate variability are represented within their learned latent spaces. In this work, we examine the internal representations of the Prithvi WxC weather–climate foundation model to determine whether they capture key modes of subseasonal monsoon dynamics. We extract regional latent representations over South Asia and the Indian Ocean from both encoder and decoder layers and use temporal sequence models, including LSTM and Transformer-based architectures, to reconstruct the Monsoon Intraseasonal Oscillation (MISO), Madden–Julian Oscillation (MJO), and Indian Ocean Dipole (IOD). By comparing different temporal input windows and latent representations, we investigate how these climate signals are organized within the model. Our analysis shows that Prithvi WxC contains meaningful information about subseasonal atmospheric variability, providing insight into the physical information learned by climate foundation models and their potential use in subseasonal climate analysis and prediction.