Emulators learn the ocean, cyclones and climate thresholds

Machine learning emulators for atmospheric processes changed weather forecasting within a few years. Ocean versions developed more slowly. The reason is visible in the physics. In the atmosphere, most energy sits in relatively large-scale phenomena. In the ocean, mesoscale eddies about an order of magnitude smaller drive the dynamics. Add convoluted coastlines, narrow straits and ice-covered seas, which atmospheric models do not have to worry about as much. Ocean simulations therefore run on locally refined or fully unstructured grids, while data models were trained on plain latitude-longitude grids.

The ocean on its own grid

Work on HClimRep-Ocean proposes an emulator that operates directly on the native unstructured grid of the FESOM2 model. The team trained it on a 209-year control integration of AWI-CM3 and runs it without atmospheric forcing. It receives the state of the atmosphere only at the start, isolating the predictability carried by the ocean itself. Skill depends heavily on the field. For currents, the emulator beats all reference models at a 30-day forecast. For temperature and salinity, the damped persistence anomaly remains better. The authors explain this physically: current variability is largely geostrophic and internally generated, while fluctuations of surface temperature and salinity are driven by weather. On the OceanBench benchmark, the variant trained on reanalysis achieved the lowest RMSE against the GLORYS reanalysis of all the systems evaluated.

Threshold scenarios at high resolution

The second set of tools assesses the effects of climate tipping points. The ClimTip-GML dataset is the first globally bias-corrected and downscaled dataset for evaluating the impact of large-scale transition scenarios. It covers eight key variables at 0.25 degree resolution from three general circulation models, CESM1-CAM5, HadGEM3-GC31-MM and MPI-ESM1-2-HR, plus century-long simulations for pre-industrial and historical conditions, and 2 degrees Celsius warming scenarios with and without a tipping transition of the Atlantic overturning circulation or the Amazon forest. Generative machine learning methods trained on reanalysis perform the bias correction and downscaling, preserving physical consistency in space, in time and across all eight variables. Validation showed a clear reduction of biases, improved small-scale variability and multivariate correlations, and a consistent long-term response to external forcing.

Physics as a constraint

The third paper deals with tropical cyclones and rapid intensification, which remains one of the hardest parts of forecasting. The authors combine two approaches. Fully physical numerical models can describe the processes responsible for intensification, but they are costly. Purely data-driven models lack physical explanation. Their hybrid, FAST-ML, uses a two-stream neural parameterization that takes three-dimensional ERA5 fields to diagnose ventilation control: the environmental wind shear and the mid-tropospheric entropy deficit. The parameters are optimized end to end through the differentiable intensity model FAST, so the evolution of the storm stays subject to the laws of thermodynamics. Against its physical reference, the model reduces the CRPS score by about 31 percent at a 60-hour horizon and nearly halves the false alarm rate for rapid intensification without losing detection sensitivity. In a hundred-member ensemble configuration, it produces forecasts comparable to FNV3 for the storms studied, and zero-shot tests in the eastern Pacific suggest portability between basins.

The shared conclusion of these papers is methodological: an emulator does not have to be universal to be useful, but it must be evaluated field by field. A model that wins on currents loses on temperature. That is not a failure, only information about where its predictability ends.