Title: Samudra 2: Scaling Ocean Emulators across Resolutions

Abstract:

While Ocean General Circulation Models (OGCMs) are critical to climate science, their high computational costs restrict the size of ensembles and the range of forcing scenarios that can be tested. Although neural emulators offer potential speedups of several orders of magnitude, current models have struggled to simultaneously achieve fine spatial resolution and multi-year autoregressive simulations. Samudra, the inaugural autoregressive neural ocean emulator capable of generating multi-decade global rollouts, operates at a $1^\circ$ resolution but suffers from two distinct long-horizon failure modes: imprinting artifacts, where velocity patterns erroneously influence deep-ocean fields, and variance collapse, characterized by a loss of temporal variability.

To address these limitations, we introduce Samudra 2. This upgraded model features a broader U-Net backbone incorporating modified ConvNeXt-style blocks with a diminished block-internal expansion factor. Additionally, it employs a dynamic loss function that reweights output channels based on prediction errors, thereby reinforcing gradients for deep-ocean fields that evolve slowly. At a $1^\circ$ resolution, Samudra 2 boosts the upper-ocean global-mean temperature $R^2$ from 0.56 to 0.87 and decreases deep-ocean temperature error by approximately seven times.

The architecture demonstrates scalability, supporting $1/2^\circ$ and $1/4^\circ$ resolutions over autoregressive rollouts of roughly eight years, successfully recovering features such as mesoscale eddies and sharp western boundary currents. By operating on a single GPU, Samudra 2 facilitates larger ensembles for research into ocean heat uptake, climate variability, and sea-level projections. Code, documentation, and benchmark resources are available at https://openathena.ai/Ocean_Emulator/.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC