An ocean-wave model tracks energy across waves with different frequencies and directions. Calculating how interacting waves redistribute that energy can be costly. NLML learns this exchange and supplies it to WAVEWATCH III while the surrounding physical model continues to run.[1]
Researchers trained it on a 2010 simulation, then tested a separate 2020 year. Compared with the usual shortcut, it reduced the global mean root-mean-square error in significant wave height by about 60%, measured against the exact interaction solver.[1]
A separate benchmark simulated one month using 64 CPU cores and four Nvidia A100 GPUs. With reduced-precision GPU arithmetic, the full model took 4% longer than the shortcut version.[1]
That measures closeness to a reference calculation and computing cost. The study does not establish a 60% improvement in operational wave forecasts.[1]
Sources & further reading
- NLML: A Deep Neural Network Emulator for the Exact Nonlinear Interactions in a Wind Wave Model Journal of Geophysical Research: Machine Learning and Computation · 23 Jan 2026 · Accessed 20 Oct 2026