Toward Coupled Coastal Wind–Wave–Current Analysis: Graph Localization and Diffusion-Learned Priors
Malaquias Peña and Panagiotis Mitsopoulos
University of Connecticut
7 October, 2026, 2-3pm
Abstract:
Surface winds, waves, and ocean currents are observed by different instruments and evolve on different time and spatial scales, yet their uncertainties are coupled. For coastal analysis, this creates two practical challenges: observations must influence the appropriate physical pathways, and background ensembles must represent dependencies across components. This seminar connects ongoing wind--wave data assimilation work for the NOAA 3D Real-Time Mesoscale Analysis with a proposed extension to surface currents. We will first discuss lessons from satellite wind and wave assimilation and idealized experiments with water-path localization, which accounts for coastlines when defining oceanic observation influence. We will then present a framework for learning joint wind--wave--current transition uncertainty with a conditional diffusion model and using the generated ensembles within a local ensemble transform Kalman filter. The framework retains an explicit forecast--analysis cycle and distinguishes learned background generation from the observation update. A coastal toy model provides a controlled setting for comparing diffusion and Gaussian priors, testing the value of current observations, and examining whether coupled-increment diagnostics track actual analysis and forecast quality. The evaluation emphasizes observation gaps, ensemble calibration, member pairing, and computational cost. Finally, I will outline a path to an Atlantic regional application using RegMOM6 backgrounds, high-frequency radar currents, and overlapping historical winds and waves, including the limits imposed by sampling cadence and data dependence.