For complex observation models (non-smooth, many-to-one, or simulator-defined), energy-based ensemble filtering with learned controlled flows outperforms traditional Kalman-type methods while remaining comparable on standard problems.
This paper introduces a new approach to data assimilation—estimating system states from model predictions and observations—that handles complex observation mechanisms like non-smooth functions or simulator-based measurements. The method uses controlled flows and energy-based updates to learn how observations should modify predictions, without requiring traditional likelihood structures.