Data-Driven Stochastic Parameterization for Unresolved Ocean Dynamics

Overview & Synopsis

Ocean models cannot resolve the full range of scales, so unresolved motions — above all mesoscale and submesoscale eddies, which set the large-scale currents, tracer transport, mixing and intrinsic variability — have to be parameterized. This talk presents a series of collaborative works on data-driven stochastic parameterization built on the stochastic transport framework, in which unresolved fluctuations enter as random perturbations of the flow map, preserving the structural properties of the fluid equations while giving a physically interpretable account of unresolved energy transfers. The noise itself is learned from high-resolution simulations: POD calibrates the dominant spatial structures of unresolved eddies in a multilayer quasi-geostrophic model, improving the jet and the low-frequency variability of coarse runs, while DMD goes beyond a stationary covariance by separating slowly evolving coherent components from faster decorrelating ones, improving variability and kinetic-energy backscatter. Both were then extended and validated in NEMO, with physical constraints — notably projecting the fluctuations along isopycnal surfaces to limit spurious diapycnal mixing — and with a hybrid scheme in which DMD learns the vertically integrated component from satellite sea-surface height while the remaining vertical fluctuations come from model-based noise. The conclusion is that unresolved eddy effects should not be modelled as purely data-driven fluctuations decoupled from the resolved flow: hybrid data- and model-driven parameterizations, learning structure from data while keeping a meaningful dependence on the simulated state, are the more effective route to uncertainty-aware operational ocean models.