AI News

DU-NO is a multiscale neural operator designed to reduce the cost of phase-resolving wave simulation, whose accuracy is useful for nearshore forecasting but traditionally makes large ensembles, uncertainty quantification, and real-time warnings impractical. The architecture adds lightweight convolutional U-Net branches only at the shallowest encoder and decoder levels, where fine-grid high-wavenumber information is present, while keeping coarse levels purely spectral. A depth-decaying mode schedule limits the model to 3.64 million parameters, compared with tens of millions for the U-FNO baseline. On the publicly released FUNWAVE-TVD benchmark, DU-NO achieved the best autoregressive rollout error among six identically trained architectures, improving on U-FNO by 14.9% with 10.8 times fewer parameters. Frequency-band analysis found the improvement across all bands, including high wavenumbers where truncated-spectral operators performed poorly. Parameter-matched controls further showed that the gain was architectural: the strongest baseline at roughly 3.6 million parameters still lagged DU-NO by 28.6%. On 2D Navier-Stokes, DU-NO matched the strongest baselines, and it outperformed them clearly on shallow-water rollouts in PDEBench. Code, trained models, and evaluation artifacts are publicly available.