ForcingDAS: Unified and Robust Data Assimilation via Diffusion Forcing
ForcingDAS: Unified and Robust Data Assimilation via Diffusion Forcing
Authors: Yixuan Jia, Siyi Chen, Yida Pan, Xiao Li, Lianghe Shi, Chanyong Jung, Haijie Yuan, Ismail Alkhouri, Yue Wu, Saiprasad Ravishankar, Jeffrey A. Fessler, Qing Qu Date: 2026-05-14 Paper ID: openalex:2605.14285
Summary
ForcingDAS is a robust data assimilation framework that addresses error accumulation in dynamical systems by learning a joint-trajectory prior rather than iterative frame-to-frame transitions. By leveraging Diffusion Forcing with frame-specific noise levels, the method effectively captures long-horizon temporal dependencies in non-Markovian scenarios. Crucially, it provides a unified interface for filtering and smoothing, allowing users to select tasks at inference time without retraining. Evaluation across multiple scientific benchmarks demonstrates that a single model can outperform regime-specific learned and classical baselines.
Key Contributions
- Introduces ForcingDAS, a unified data assimilation framework that replaces frame-to-frame transition models with joint-trajectory priors via Diffusion Forcing.
- Enables a single trained model to perform both filtering (nowcasting) and smoothing (reanalysis) tasks by adjusting the inference schedule without retraining.
- Achieves superior robustness against error accumulation in non-Markovian observational settings, outperforming specialized baselines on 2D Navier-Stokes and global atmospheric state estimation benchmarks.
Open Questions & Future Work
Key Concepts
- diffusion-forcing: A training technique that assigns independent noise levels to individual frames, allowing models to learn joint-trajectory priors instead of traditional frame-to-frame transitions.
Archivist Review
The paper introduces a significant shift in data assimilation by moving from frame-to-frame transitions to trajectory-based diffusion priors. ‘Diffusion Forcing’ is approved as a foundational concept. The two open questions address critical bottlenecks in existing assimilation paradigms—specifically architectural limitations in backward information flow and the need for physical consistency in data-driven models—which are central challenges in time-series and scientific machine learning.
Approved Concepts
- Diffusion Forcing: Central mechanism enabling the framework to learn joint-trajectory priors and avoid error accumulation in dynamical systems.
Approved Open Questions
- Causal-only Smoothing Bottleneck: This is a fundamental architectural limitation that restricts the efficacy of information propagation from future observations to past states, which is critical for high-performance smoothing and reanalysis.
- Physics-constrained Assimilation Integration: Integrating physics constraints is crucial for ensuring the physical realism and reliability of assimilated state estimates in scientific and climate applications.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2605.14285
- paper_id
- 2605.14285
- paper_source
- openalex
- domain
- time-series
- tags
-
- architectures
-
- datasets
-
- concept_slugs
- diffusion-forcing
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-05-17T06:09:00Z