Flow-Matched Data Assimilation for Historical Ocean-Current Reconstruction

Historical ocean-current products often trade spatial resolution against temporal resolution, limiting analyses that require both fine spatial structure and subdaily variability. We present a generative data-assimilation framework that combines complementary historical products with a high-resolution numerical-model prior. A conditional flow-matching model jointly generates 24-hour current fields from numerical-model conditioning, and FlowDAS guides these fields using a daily, spatially detailed GLORYS12V1 constraint and a higher-frequency, spatially coarser GlobCurrent constraint. On a held-out test set, FlowDAS reduces mean-squared error (MSE) against GLORYS12V1 from 0.4405 to 0.1081 (75\%) and against GlobCurrent from 0.7786 to 0.1364 (82\%). Against independent drifter-derived measurements excluded from the assimilation likelihood, MSE decreases from 1.0848 to 0.5706 (47\%) and vector correlation increases from 0.7806 to 0.8309. Spatial, temporal, and kinetic-energy diagnostics show that the method transforms a numerical-model-aligned prior into a reanalysis-like posterior reconstruction while retaining the model grid and cadence.