Samudra 2: Scaling Ocean Emulators across Resolutions
Researchers have developed Samudra 2, an advanced neural emulator for ocean circulation models that significantly improves accuracy and speed. This new model addresses limitations of its predecessor, such as variance collapse and imprinting artifacts, by employing a wider U-Net backbone and a dynamic loss function. Samudra 2 achieves higher accuracy in predicting ocean temperatures and can simulate finer resolutions over longer time scales, enabling more extensive climate studies. AI
IMPACT Enhances climate modeling capabilities, enabling larger ensembles for sea-level and heat uptake projections.