Speaker
Description
Machine learning is reshaping how we simulate the Earth system: data-driven weather models now rival operational forecasts at a fraction of their cost. Whether the same holds for the ocean is far less clear, as it is a slowly evolving, strongly forced component whose simulations have to stay stable far beyond weather forecast horizons. At GEOMAR we follow two complementary routes to machine-learning based ocean and sea-ice modeling; this first of two talks presents the route that replaces the numerical model entirely with a data-driven emulator.
A key insight from recent advances in numerical ocean modeling is that small-scale processes are not merely a detail: their representation is necessary even for a correct simulation of large-scale features such as the Gulf Stream path. This motivates training on kilometer-scale simulation data rather than coarser reanalysis: the training signal must contain the small-scale variability that shapes large-scale structure, even if the emulator itself operates at reduced resolution. Two tensions then arise: capturing the stochastic imprint of these processes to produce realistic mesoscale variability, while maintaining rollout stability over months to years.
We present RánCast, a global 3D ocean emulator trained on 30 years of ocean data from the coupled FESOM–IFS nextGEMS simulation, regridded from kilometer-scale to 1.5° resolution. The architecture adapts ArchesWeather to the ocean domain, predicting sea surface height, potential temperature, salinity and horizontal velocities as daily snapshots. We evaluate two additions aimed at improving multi-month to multi-year rollout stability. First, atmospheric surface forcing (2-m temperature, surface pressure, precipitation, 10-m winds) from the IFS component of the same simulation is re-injected as encoder input at every autoregressive step, providing a physically grounded external constraint that anchors the ocean trajectory and preserves mesoscale variability where forcing dominates. Second, a learnable residual scaling coefficient per variable and depth level replaces the fixed skip connection, letting the model adapt the persistence–increment balance to the dynamics at each depth.