DU-NO: A Parameter-Efficient Double U-Shaped Neural Operator for Phase-Resolving Wave Modeling

Forum topic · 小凯 · 2026-09-15

Summary

DU-NO (Double U-shaped Neural Operator) is a multiscale U-shaped spectral operator designed to emulate phase-resolving nearshore wave models such as FUNWAVE-TVD at a fraction of the computational cost. The architecture attaches lightweight convolutional U-Net branches only at the two shallowest encoder and decoder levels, where fine-grid, high-wavenumber content resides, while coarser band-limited levels remain purely spectral. A depth-decaying mode schedule keeps the model at 3.64 million parameters, roughly an order of magnitude smaller than the U-FNO baseline. On a public FUNWAVE-TVD benchmark, DU-NO achieved the best autoregressive rollout error among six identically trained architectures, improving over U-FNO by 14.9% with 10.8x fewer parameters, with gains holding across all frequency bands including high-wavenumber bands where truncated spectral operators collapse. Parameter-matched controls show the best baseline still trails DU-NO by 28.6% at equal 3.6M parameter budgets, confirming architectural benefits. DU-NO also matches the strongest baseline on 2D Navier-Stokes and clearly wins on PDEBench shallow-water rollouts.

Overview

This post introduces the paper DU-NO: A Parameter-Efficient Double U-Shaped Neural Operator for Phase-Resolving Wave Modeling (arXiv: 2609.12115) by Enrique Hernandez Noguera, Md Meftahul Ferdaus, Nathan Cooper, Elias Ioup, and Mahdi Abdelguerfi.

Key Points

Motivation: Phase-resolving wave models like FUNWAVE-TVD are the accuracy standard for nearshore dynamics (resolving shoaling, refraction, and breaking of individual waves), but their cost makes them impractical for operational ensembles, uncertainty quantification, and real-time warning.
Problem with existing neural operators: Hybrid spectral-convolutional operators that achieve high fidelity on wave-dominated fields (e.g., U-FNO, the strongest baseline in the study) require tens of millions of parameters.
Architecture: DU-NO is a multiscale U-shaped spectral operator with lightweight convolutional U-Net branches attached only at the two shallowest encoder and decoder levels. This follows a sampling argument: high-wavenumber content exists only on fine grids, so the local full-band pathway belongs there, while coarse band-limited levels remain purely spectral.
Efficiency: A depth-decaying mode schedule keeps the model at 3.64 million parameters, an order of magnitude below U-FNO.

Results

Best autoregressive rollout error among six identically trained architectures on a public FUNWAVE-TVD benchmark, improving over U-FNO by 14.9% with 10.8x fewer parameters.
Frequency-band analysis shows the gains hold across all bands, including high-wavenumber bands where truncated spectral operators collapse.
Parameter-matched control: after rescaling the best baseline to the same 3.6M parameter budget, it still trails DU-NO by 28.6%, confirming the gains are architectural.
Generalization beyond nearshore waves: DU-NO ties the strongest baseline on 2D Navier-Stokes and clearly wins on PDEBench shallow-water rollouts.

References

Paper: <https://arxiv.org/abs/2609.12115>