cs.LGNov 18, 2026

QuantAtlas-OM: Taming Neural Spectral Bias in Long-Horizon Spatiotemporal Ocean Dynamics

Authors: Hao Wu, Yuan Gao, Tong Wu, Fan Xu, Fan Zhang, Jinhan Li, Kun Wang, Zhenyu Ma, 

Organizations: THU · Ubiquant Investment · IQuest Research · USTC · 

Abstract

Autoregressive neural networks exhibit spectral bias in long-horizon spatiotemporal prediction. As rollouts progress, high-frequency details decay, blurring the predicted fields and destabilizing long-range forecasts. Ocean dynamics offer a particularly stringent test because kinetic energy is concentrated in mesoscale eddies and fronts spanning tens of kilometers. This degradation stems from coupling large-scale phase evolution and small-scale energy retention in a shared feature space. Evolution on a coarse grid smooths out high-wavenumber eddies, whereas evolution on a fine grid amplifies small-scale phase errors and triggers rollout instability. We introduce QuantAtlas-OM (QuantAtlas Ocean Model), a scale-spectral decoupling framework that confines deep temporal evolution to a low-wavenumber latent manifold for long-term stability while routing full-resolution eddy and front residuals through an outer bypass to the decoder. Across physical regimes, the evolution core pairs non-local attention with resolution-invariant dilated spatial gating. Across three western boundary currents, QuantAtlas-OM retains 43.7%43.7\% to 86.5%86.5\% of mesoscale spectral energy along the k−3k^{-3} cascade at Day 120120, sustains anomaly correlation coefficients of 0.810.81 to 0.880.88 at Day 6060, and reaches 0.660.66 at Day 120120 in the Kuroshio (2.5×2.5\times to 7.7×7.7\times the best baselines). Trained at 0.25∘0.25^\circ, QuantAtlas-OM transfers zero-shot to a global 0.125∘0.125^\circ grid (0.810.81 Day-1010 spatial correlation) and leads global 0.25∘0.25^\circ, 2323-level 3D ocean simulations (120120 days) and coupled ocean-atmosphere forecasts (6060 days), ranking first on 1515 of 1616 metrics and cutting 6060- to 120120-day salinity RMSE by 13%13\% to 27%27\% across global ocean basins.

Figures & tables

Explore similar work

CardsList
  1. Samudra 2: Scaling Ocean Emulators across Resolutions

    May 24, 2026Yuan Yuan, Jesse Rusak, Alexander Merose +5Neural Surrogate ModelingClimate Science

  2. Neptune: An AI model for Global Ocean Subseasonal Prediction

    Sep 8, 2026Davide Donno, Italo Epicoco, Massimo Cafaro +8Neural Surrogate ModelingClimate Modeling

  3. KFTD: Koopman-Fourier Time-Differentiable Network for Continuous Ocean Spatiotemporal Forecasting

    Jun 6, 2026Qinghui Chen, Zekai Zhang, Hailong Liu +2Koopman Operator LearningPhysics-Informed ML