Appa: Bending Weather Dynamics with Latent Diffusion Models for Global Data Assimilation

We introduce a score-based data assimilation framework powered by a 565M-parameter latent diffusion model trained on ERA5 reanalysis data. Our model generates global atmospheric trajectories at 0.25 degree resolution and 1-hour intervals, and supports conditioning on arbitrary observations to infer plausible trajectories, without retraining. This unified probabilistic approach handles reanalysis, filtering, and forecasting tasks while maintaining physical consistency.

In Machine Learning and the Physical Sciences (ML4PS) Workshop @ NeurIPS 2025.