Samudra 2

A neural ocean emulator from NYU's M2LInES group (first author Yuan Yuan, senior author Laure Zanna) with Open Athena engineers Jesse Rusak and Alexander Merose as second and third authors, plus Princeton's Alistair Adcroft. Ocean general circulation models are the expensive part of climate simulation, which limits how many scenarios and ensemble members researchers can afford. The original Samudra was the first autoregressive neural ocean emulator to produce multi-decade global rollouts, but it ran only at 1° and drifted over long horizons in two ways: variance collapse, where temporal variability fades, and imprinting artifacts, where velocity patterns leak into deep-ocean fields.

Samudra 2 fixes both with a wider U-Net backbone of modified ConvNeXt-style blocks (stage widths raised from [200, 250, 300, 400] to [280, 380, 480, 520] while the block-internal expansion factor drops from 4 to 2) and a dynamic variance-weighted loss that updates per-channel MSE weights online from an exponential moving average of inverse prediction error, so slow-moving deep-ocean fields keep a usable gradient. At 1° it raises upper-ocean global-mean temperature R² from 0.56 to 0.87 and cuts deep-ocean temperature error roughly sevenfold. The same architecture scales to 1/2° and 1/4° over about eight-year free-running rollouts, where mesoscale eddies and sharp western boundary currents emerge, and it runs on a single GPU. The emulator maps two consecutive ocean states plus atmospheric forcing to the next two, over 77 prognostic channels (temperature, salinity and both velocity components across 19 depth levels, plus sea surface height), trained on GFDL OM4 output. Open Athena's write-up reports extending it to a dataset 16 times larger on the same hardware budget and cutting training at the original resolution from four days to four hours. Code is Apache 2.0; weights are on HuggingFace under CC BY 4.0.