ForcingDAS unifies data assimilation with diffusion models, beating specialized baselines

Data assimilation (DA) traditionally splits into filtering (real-time nowcasting) and smoothing (retrospective reanalysis), each requiring separate models. These often fail under non-Markovian observations (partial slices of a latent state) by accumulating errors over long horizons. To solve both issues, Yixuan Jia and colleagues from multiple institutions propose ForcingDAS, a unified framework built on the recently introduced Diffusion Forcing technique. Instead of frame-to-frame transitions, ForcingDAS learns a joint-trajectory prior with independent noise levels per frame, enabling it to capture long-range temporal dependencies and avoid error accumulation.

Critically, the same trained model can operate across the full filtering-to-smoothing spectrum purely by adjusting the inference schedule—no retraining needed. The authors demonstrate this on three challenging benchmarks: 2D Navier-Stokes vorticity, precipitation nowcasting, and global atmospheric state estimation. In every case, ForcingDAS is competitive with or outperforms both learned models and classical methods that are specialized for a single regime, with the largest gains on real-world weather data. This work could significantly streamline weather forecasting and climate modeling pipelines.