Flow matching for posterior sampling in variational data assimilation
OpenVariational data assimilation gives a mode of the posterior; generative models can give samples. This project investigates bridging variational assimilation with flow matching so that the assimilation system produces genuine posterior samples — the direction I presented at the EnKF Workshop 2026.
Literature & running notes
EnKF Workshop 2026 talk — Bridging Variational DA and Flow Matching for Posterior Sampling
My talk sketching the connection; the project is to develop this into a working method on a quasi-geostrophic testbed and compare against ensemble Kalman baselines.
References
https://arxiv.org/abs/2510.02952
Concepts for interpretability, physical guidance in models
Concept bottleneck models (CBMs) offer a middle ground by routing predictions through an intermediate layer of interpretable concepts. They structurally embed physical reasoning into the model without over-constraining it. Tp recover the mechanistic drivers of marine heatwaves (MHWs) through prescribed concepts and a free concept that captures residual structure and regularizes prescription. By combining mixed supervision with an ensemble of CBMs, OceanCBM balances physical constraint, flexibility, and mechanistic interpretability. https://arxiv.org/pdf/2605.12639