Physically Consistent Sampling For Ocean Model Initialization

Accurate simulation of ocean variability and climate response relies on an initialization phase called "spin-up" during which the ocean model reaches equilibrium under an applied forcing.This process comes at a high computational cost that can amount to over 1000 simulation years. Recent advances in deep generative models as climate emulators offer promising acceleration opportunities through their efficiency and ability to capture and generate complex spatio-temporal patterns. However, existing generative models often produce physically inconsistent results due to incomplete representation of underlying physical laws. In this work, we leverage on recent climate emulators to accelerate the initialization phase of ocean models. We introduce physically meaningful constraints on vertical stratification that guide sampling toward physically coherent results. Experiments on idealized ocean simulations demonstrate successful enforcement of vertical stratification.