TheMachine Press

A daily newspaper for the age of artificial intelligence.

Morning editionPermanent story

research

A Wave Model Kept the Detail With One-Tenth the Parameters

DU-NO used 3.64 million parameters, improved rollout error 14.9% over U-FNO and preserved high-frequency wave structure.

Published Updated Story ID: mp-2026-09-14-004
Read the complete editionStory JSON

Summary

DU-NO used 3.64 million parameters, improved rollout error 14.9% over U-FNO and preserved high-frequency wave structure.

Phase-resolving coastal models are accurate but too expensive for many ensembles and real-time forecasts. DU-NO places lightweight convolutional U-Net branches only on the fine encoder and decoder levels where high-wavenumber content exists, leaving coarser levels spectral. The resulting 3.64-million-parameter neural operator used 10.8 times fewer parameters than U-FNO while improving autoregressive rollout error by 14.9% on the authors’ FUNWAVE-TVD benchmark. A parameter-matched baseline still trailed by 28.6%, and tests extended to Navier–Stokes and shallow-water rollouts. Operational warning performance remains to be established.

Why it matters

DU-NO used 3.64 million parameters, improved rollout error 14.9% over U-FNO and preserved high-frequency wave structure.

Limits and context

  • DU-NO places lightweight convolutional U-Net branches only on the fine encoder and decoder levels where high-wavenumber content exists, leaving coarser levels spectral.

Key claims

  1. DU-NO used 3.64 million parameters, improved rollout error 14.9% over U-FNO and preserved high-frequency wave structure.

    Qualification: DU-NO places lightweight convolutional U-Net branches only on the fine encoder and decoder levels where high-wavenumber content exists, leaving coarser levels spectral.

    Evidence: source-2026-09-14-004

Sources

  1. arXiv preprint 2609.12115arXiv · primary research

Corrections

No corrections have been recorded for this story.