PR-Smoother: Simulator-Preserving Non-Gaussian Smoothing for Data AssimilationNew

PR-Smoother is a simulator-preserving amortized smoother designed for physical data assimilation workflows. It addresses the challenge of representing non-Gaussian posteriors over physical state variables while scaling to high-dimensional simulators. The method trains from observation windows alone and remains compatible with calibration of the prescribed simulator, making it suitable for earth science and geophysical applications requiring accurate state estimation under uncertainty.