DU-NO Model Released: Parameter-Efficient Double U-Shaped Neural Operator Revolutionizes Nearshore Wave Modeling

Summary:DU-NO (Double U-shaped Neural Operator) is a novel multiscale U-shaped spectral operator that attaches lightweight convolutional U-Net branches at its two shallowest encoder and decoder levels. With only 3.64 million parameters, it is an order of magnitude smaller than the strongest baseline, U-FNO, while achieving a 14.9% reduction in autoregressive rollout error on the FUNWAVE-TVD benchmark. The model also demonstrates superior performance across all frequency bands, including high-wavenumber

---

Key Breakthroughs

• Efficient Parameterization: DU-NO incorporates lightweight convolutional U-Net branches at its shallowest encoder and decoder levels, reducing the number of parameters to just 3.64 million, an order of magnitude smaller than the strongest baseline, U-FNO.
• Multiscale U-shaped Spectral Operator: The model employs a multiscale U-shaped spectral operator architecture, guided by a sampling argument that restricts high-wavenumber content to fine grids while keeping coarse, band-limited levels purely spectral.
• Significant Performance Gains: On the FUNWAVE-TVD benchmark, DU-NO achieves a 14.9% reduction in autoregressive rollout error compared to U-FNO, with superior performance across all frequency bands, including high-wavenumber regions where truncated-spectral operators typically fail.
• Cross-Domain Applicability: DU-NO not only excels in nearshore wave modeling but also demonstrates strong performance in 2D Navier-Stokes and PDEBench shallow-water simulations.

Technical Highlights

1. Depth-Decaying Mode Schedule: The depth-decaying mode schedule mechanism allows DU-NO to maintain a compact parameter count of 3.64 million, significantly reducing computational costs.
2. Lightweight Convolutional U-Net Branches: The addition of lightweight convolutional U-Net branches at the shallowest levels enhances the model's ability to capture local features.
3. Frequency Band Analysis: DU-NO performs exceptionally well across all frequency bands, particularly in high-wavenumber regions, outperforming traditional truncated-spectral operators.

Industry Impact

The release of DU-NO provides a new technical pathway for nearshore wave modeling and complex physical simulation fields. Its efficient parameterization and superior performance make it highly promising for applications in real-time warning, ensemble simulation, and uncertainty quantification. Additionally, DU-NO's cross-domain applicability extends its potential use in areas such as fluid dynamics and meteorological forecasting.

Recommendations for Developers

• Model Application: Developers are encouraged to apply DU-NO to nearshore wave modeling and complex physical simulation tasks to validate its performance in real-world scenarios.
• Parameter Adjustment: Depending on the specific application, developers can further adjust the depth-decaying mode schedule mechanism to optimize model performance.
• Cross-Domain Exploration: Researchers are encouraged to explore DU-NO's applications in fluid dynamics, meteorological forecasting, and other domains to expand its range of use.

---

Source: ArXiv AI (cs.AI) (2026-09-14)

— END —

Tags: #DU-NO #Neural Operator #Nearshore Wave Modeling #Parameter Efficiency #Multiscale Architecture