A compact artificial-intelligence model advanced a coastal-wave simulation 173 times faster than the numerical solver used to create its training data. Across 16 held-out cases, it also produced less accumulated error than the strongest rival trained under the same conditions. The LSU New Orleans and U.S. Naval Research Laboratory study was posted as an unreviewed preprint on September 10.
Coastal models calculate how waves bend, grow and break as they enter shallow water. That detail helps researchers study flooding and erosion, but repeatedly solving the underlying equations is too slow for large forecast ensembles. The new model, called DU-NO, learns the change from one surface-elevation map to the next and feeds each prediction back into the following step.
DU-NO divides the calculation by scale. Fourier layers carry broad wave patterns through the model, while small convolutional branches preserve sharp crests and breaking-wave detail near the highest-resolution stages. Deeper stages work on coarser representations. This arrangement concentrates memory and computation where local detail is most likely to disappear.
The team generated 81 simulations over measured seabed topography at Duck, North Carolina. Wave height, frequency, direction and water depth were varied systematically; 57 cases were used for training and 16 for the final test. Six neural architectures received the same ten input frames and predicted the following ten without seeing the correct intermediate results.
DU-NO recorded a relative error of 0.0946, compared with 0.1111 for the next-best U-FNO model, where lower is better. It used about one-tenth as many parameters. On the same 16-core CPU, one DU-NO step took 21.5 milliseconds and delivered the reported 173-fold speedup over the coastal solver. The gain therefore combined faster calculation with a modest improvement over the other learned model, not greater accuracy than the original physics solver.
The experiment covers one coastline, regular two-dimensional grids and wave fields produced by a simulator rather than measured at sea. It predicts surface height but not the horizontal velocities needed to estimate currents. Tests against observations from several coasts will determine whether the speed advantage remains useful when real waves depart from the training simulations.
Sources
- Enrique Hernandez Noguera and colleagues, arXiv, September 10, 2026. Abstract, authorship, submission date and principal findings.
- Full paper. Model architecture, benchmark design, comparisons, timing and limitations.
- Lead image: JessicaVz, Wikimedia Commons, licensed under CC BY 4.0. The photograph is illustrative and does not show the reported benchmark.
