DU-NO architecture combines spectral and convolutional operators: coastal wave simulation with 3.6 million parameters

A research team that includes Enrique Hernandez Novgira, Mohamed Miftah Al-Firdous, Nathan Cooper, Elias Youb and Mahdi Abdelkharfi presented a paper that introduces a new neural architecture called DU-NO for dual-form neural network operators, intended for simulating near-shore ocean wave motion with accuracy based on physical phase decomposition. The paper addresses an existing computational dilemma in hydrodynamic modeling, where traditional physical models such as FUNWAVE-TVD constitute the primary benchmark for accurately capturing wave refraction phenomena, their phases and energy dissipation, but their high computational cost prevents their use in real-time forecasting systems, early-warning deployment, or the construction of iterative probabilistic models designed to estimate uncertainty margins.

The architectural innovation of the DU-NO model came to untie the existing link between neural operator accuracy and the inflation of their software parameter count.While previous hybrid models such as U-FNO rely on tens of millions of parameters to regulate the behavior of complex wave fields, the new model starts from a mathematical hypothesis linked to the nature of sampling, confining the lightweight convolutional branches to only the top two levels of the encoder and decoder, based on the premise that high-frequency wave components appear only in fine spatial grids, whereas the deeper levels remain purely spectral according to a schedule that reduces patterns linearly as depth increases. This geometric allocation limits the model size to just 3.64 million parameters, roughly 10.8 times smaller than the competing U-FNO model.

Experimental results published on the FUNWAVE-TVD evaluation metric showed that the new architecture outperforms in self-regressive rollout by 14.9 percent compared with the U-FNO model, with this improvement extending across all frequency bands, including high-wave-number ranges where truncated spectral operators typically fail. Comparative calibration tests confirmed that the gains stem from the architectural design itself, as the best competing model lagged by 28.6 percent when its budget was constrained to the same computational ceiling of 3.6 million parameters. The model also demonstrated performance on par with the strongest models on two-dimensional Navier-Stokes equations and a clear advantage in shallow-water simulation within the PDEBench benchmark.

I see this reduction in parameter burden as a practical transition for marine engineering teams and coastal infrastructure departments in the Arabian Gulf, the Red Sea and the eastern Mediterranean coasts.Shifting wave simulation from supercomputing servers to an architecture that requires only about 3.6 million parameters enables port operators and marine energy platform managers in the region to run hydrodynamic forecasts locally without draining cloud-computing budgets. As a coastal engineer or planner you can simulate hundreds of concurrent wave scenarios to test the safety of breakwaters and piers before marine storms with full physical accuracy and instantaneous response time.

With the source code, models and data made publicly available, the next step for technical teams in the region becomes feeding this engineering framework with locally sourced coastal depth data for bays and harbors, and leveraging the reduced inference cost to integrate real-time simulation into maritime control rooms and environmental systems without the need to purchase expensive central computing.