WorksComputer Science
QuantAtlas-OM: Taming Neural Spectral Bias in Long-Horizon Spatiotemporal Ocean Dynamics
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 to of mesoscale spectral energy along the cascade at Day , sustains anomaly correlation coefficients of to at Day , and reaches at Day in the Kuroshio ( to the best baselines). Trained at , QuantAtlas-OM transfers zero-shot to a global grid ( Day- spatial correlation) and leads global , -level 3D ocean simulations ( days) and coupled ocean-atmosphere forecasts ( days), ranking first on of metrics and cutting - to -day salinity RMSE by to across global ocean basins.