Ocean eddies skip the coarse grid: learning evolution at 0.1 degrees keeps mesoscale energy and frontal gradients

Resolving ocean mesoscale eddies requires AI to learn dynamical evolution directly in a high-resolution state space, rather than advancing states on a coarse grid and restoring detail by spatial downscaling. The prior alternative evolved ocean states on a 1-degree grid and reconstructed high-resolution fields by downscaling; that route recovers large-scale patterns but fails to reproduce key mesoscale signatures. Direct learning of dynamical evolution on a 0.1-degree grid represented mesoscale variability more faithfully than 1-degree evolution followed by downscaling reconstruction.

The approach trains two emulators, one covering the full depth at 1 degree and one for 0.1-degree sea surface temperature, each advancing ocean states in free-running mode. Both emulators are used for long-term free-running simulation rather than short forecast-style stepping, making stability a precondition for the comparison. Both remained stable over long-term free-running simulations without significant drift, forming reliable baselines for the comparison.

What direct learning preserves is mesoscale energy, spectral characteristics, and frontal gradients, exactly the quantities the coarse-evolution-plus-downscaling route loses. The coarse-grid route recovers only large-scale patterns; mesoscale signals are not restored by downscaling reconstruction. The learned mesoscale statistics remained consistent across diverse hardware platforms, supporting reproducibility and portability of AI-based ocean simulation.

The comparison concerns representation of mesoscale variability along two routes: direct learning on a 0.1-degree grid and 1-degree evolution followed by downscaling reconstruction, evaluated on mesoscale energy, spectral characteristics, and frontal gradients. The two emulators are a full-depth 1-degree emulator and a 0.1-degree sea surface temperature emulator, both stable over long-term free-running simulations without significant drift. The material does not state how far these conclusions extend to other resolutions, other ocean variables, or particular regions.

The material is a summary and gives no error values, evaluation period length, or specific evaluation metrics, so the size of the gap between the two routes cannot be judged from the available information. The cross-platform consistency of mesoscale statistics supports reproducibility and portability, which is not the same as long-term stability on other platforms.