Land and Atmospheric Forecasting with a Pretrained JEPA-based Earth System Model

Data-driven Earth system models typically forecast directly in physical space. The WeatherGenerator (WeGen) instead uses a temporal Joint-Embedding Predictive Architecture (JEPA) to forecast Earth system representations in latent space. Heterogeneous land and atmospheric observations are fused into a single learned state, from which physical quantities are recovered by lightweight decoders. We evaluate this state by freezing the pretrained backbone and training task-specific decoders for four forecasting applications: precipitation, land surface temperature, Sentinel-1 radar images, and surface station variables. The resulting representations support stable ten-day autoregressive forecasts, improve land surface temperature over climatology during peak daytime hours, reconstruct meaningful high-resolution radar detail, and decode temperature, pressure, and wind at irregular station coordinates. These results demonstrate that temporal predictive representation learning provides a flexible basis for forecasting across Earth system components.