A boundary-informed UNet plus plus -ConvLSTM emulator of a SWAN-based Delft3D-FM WAVE hindcast for nearshore wave-field reconstruction

Citations

WEB OF SCIENCE

0
Citations

SCOPUS

0

초록

As sea level is increasing, fast reconstruction of high-resolution coastal wave-model fields is increasingly crucial for early warning and risk management. Physics-based wave spectral models, such as SWAN, are computationally expensive for real-time and ensemble forecasting. Meanwhile, global-scale deep-learning wave emulators require massive training data, but operate at coarse spatial and temporal resolutions, limiting their usefulness for some coastal guidance applications. To overcome this limitation, we developed a regional one-step emulator that reconstructs hourly nearshore wave fields from a Delft3D-FM WAVE model coupled with FLOW around the Korean Peninsula. Compared with global-scale emulators, our regional high-resolution framework is designed to represent bathymetry-related variations in nearshore wave fields with practical computational and memory demands. The architecture combines a UNet++ with a ConvLSTM architecture to enhance spatiotemporal representation and emulation skill. Controlled ablations show that removing spectra-derived open-boundary descriptors increases H-S RMSE by about 70%, T-m RMSE by 49%, and directional cRMSE by 39%. Boundary descriptors are the most influential input group among the tested ablations. Depth-averaged currents and nested UNet++ deep supervision are added to improve emulation performance. The high-resolution emulator first captures the nearshore spatial variability caused by bathymetry. It then reconstructs the parent-model fields with low domain-wide error for held-out test periods. Because the emulator is trained to reproduce the parent model, it does not correct parent-model biases. The emulator and the parent model under-estimated the typhoon wave peaks at the buoys. The emulator, however, reconstructs the wave component on a single desktop GPU about 210 to 240 times faster than the Delft3D-FM WAVE run computes it. This speedup covers the wave component only, because the emulator replaces the wave-prediction part and does not emulate the FLOW solution.

키워드

SWAN; Surrogate modeling; UNet plus plus -ConvLSTM; Deep-learning emulator; Computational acceleration; Nearshore waves; NEURAL-NETWORKS; WAVEWATCH III; SYSTEM; MODEL
제목
A boundary-informed UNet plus plus -ConvLSTM emulator of a SWAN-based Delft3D-FM WAVE hindcast for nearshore wave-field reconstruction
저자
Kim, Jinyoung; Son, Sangyoung
DOI
10.1016/j.apor.2026.105211
발행일
2026-09
유형
Article
저널명
Applied Ocean Research
권
174