Posters 1

[Poster 1-08] Conditional Flow Matching-Based Surrogate Modeling of Lagrangian Transport for Floating Marine Debris in a Tide-Dominated Coastal Region

*Myeongmo Kang (1. Department of Marine Environmental Science, Chungnam National University (Korea))

Conventional Lagrangian particle tracking requires repeated spatiotemporal interpolation and numerical integration of hydrodynamic velocity fields, resulting in substantial computational cost. This study proposes a conditional flow matching surrogate model for floating marine debris transport in a tide-dominant coastal area. The surrogate learns a continuous-time vector field from Gaussian probability paths constructed around spline-interpolated reference trajectories. Four configurations were evaluated: an unconditioned model, early-time displacement conditioning, M2 tidal phase conditioning, and multi-tidal conditioning. Performance was assessed using root mean square error (RMSE), temporal changes in spatial particle distributions, and integration time. All models reproduced the dominant oscillatory behavior and overall transport direction of the reference trajectories, while RMSE decreased progressively from the unconditioned model to the multi-tidal conditioning model. The multi-tidal model achieved the lowest RMSE at every time step, demonstrating the benefit of incorporating multiple tidal constituents. For 10,000 trajectories, the surrogate required 0.20 s for integration compared with 6.05 s for the reference simulation, corresponding to a 30.25-fold speedup. These results demonstrate that conditional flow matching can provide an efficient surrogate for Lagrangian particle tracking.