Abstract We present AtlanticWave, a spectral ocean wave model implemented in Python/PyTorch executing on consumer graphics processing unit (GPU) hardware (Apple M3 Max). Motivated by the need to predict surf conditions behind the Bahamas swell shield (a multi-scale problem that operational models at 9–25 km resolution cannot resolve), the model introduces five innovations, three of which are highlighted here: (1) equilibrium-referenced source terms (KEWL) where each term independently targets a known equilibrium rather than relying on compensating errors between matched pairs, (2) MERMAID (specifically v10), a learned approximation to the exact nonlinear transfer integral that achieves r = 0.89 correlation with exact Boltzmann integral solutions at comparable computational cost to the DIA ( r ∼ 0.5–0.7), and (3) a satellite-calibrated correction matrix that compensates for unresolved sub-grid bathymetric features by applying a persistent, direction-dependent nudge trained from altimeter observations, distinct from standard data assimilation and, to our knowledge, not previously published for spectral wave models. Pooled validation against 12 NDBC buoy stations across 4 hindcast periods (7,571 paired observations over 768 forecast-hours) yields significant wave height (Hs: the average height of the largest third of waves) RMSE = 0.435 m, correlation = 0.836, and bias = +0.069 m, with lower systematic bias than NOAA WW3 (RMSE = 0.381, correlation = 0.897, bias = +0.186 m) while running on a laptop rather than HPC infrastructure. AtlanticWave initializes from ECMWF WAM analysis fields; WAM and WW3 each run their own independent data assimilation cycles. All three run as free-running forecasts from their respective initial conditions. A triple-nested grid (Atlantic 28 km, Caribbean 5 km, Cabarete 500 m) resolves swell filtering through island passages invisible to coarser grids. A 192-h G1-only Atlantic forecast (0.5°) runs in approximately 4 min; the full triple-nested forecast (G1 0.25° + G2 + G3) completes in under 25 min.
David M. Mody (2026) studied this question.