Flow Matching for Efficient and Scalable Data Assimilation

This work introduces the ensemble flow filter (EnFF), a training-free data assimilation framework based on flow matching. EnFF uses Monte Carlo estimators for the marginal flow field and localized guidance to assimilate observations, and it leverages a flow design aligned with Bayesian DA. The method generalizes classical filters (e.g., bootstrap particle filter and ensemble Kalman filter) and, on high-dimensional benchmarks, shows improved cost–accuracy trade-offs and scalability, highlighting flow matching as a practical tool for efficient DA.