A hybrid system combined an AI weather forecast’s large-scale skill with the detailed storm simulations of a physics-based model.
Researchers tested a hybrid version of the European Centre for Medium-Range Weather Forecasts’ ensemble system. The method adjusted only the broad patterns of upper-air temperature and vorticity toward forecasts from a probabilistic AI system, while the physics-based model continued to simulate smaller-scale weather features.
Compared with the unmodified physics-based ensemble, the hybrid system extended useful forecast skill by up to two days in the tropics and about half a day outside the tropics. Tropical cyclone track errors also fell, while intensity forecasts and ensemble spread remained unchanged.
Better tracks, steady intensity
The hybrid ensemble showed substantial gains in large-scale forecast skill relative to the unmodified physics-based ensemble. The improvement extended forecast skill by up to approximately two days in the tropics and by about half a day in the extra-tropics. Improvements also appeared in several near-surface measures, including 2-metre temperature, 10-metre wind speed and precipitation, even though the method adjusted only upper-air fields.
Tropical cyclone track forecasts improved significantly. The study reports reductions in mean position errors of up to approximately one day, while tropical cyclone intensity forecasts remained unchanged. The ensemble spread, which represents the range of forecast outcomes, was not degraded. Tests also found that the hybrid system preserved small-scale variability and showed no instability or episodic forecast failures in the forecasts examined.
Combining different strengths
The results suggest that AI and physics-based forecasting systems can contribute different strengths within one ensemble. The AI forecasts supplied information about broad atmospheric patterns, including the steering flow that influences tropical cyclone movement, while the physics-based model retained its detailed representation of smaller-scale structures and its wider set of forecast outputs.
That combination improved track forecasts without worsening intensity or uncertainty estimates in this study. It provides a tested route for using machine-learned information to constrain an ensemble forecast while allowing the physics-based model to continue simulating finer-scale weather.
How the test was run
The evidence comes from forecast experiments using eight-member ensembles with 15-day forecasts. The researchers ran nudged and un-nudged forecasts initialized every day at 00 and 12 UTC during 1 July to 13 November 2024 and 1 December 2024 to 28 February 2025, comparing the hybrid system with the standard physics-based ensemble and the AI ensemble used for nudging.
The experiments used a reduced ensemble size and cover the stated forecast periods rather than all seasons or weather conditions. The results therefore show how the approach performed in these tests; the excerpts do not establish how it would perform with larger ensembles, different periods or other forecasting systems. The study also notes that applying the method increases the computational cost of the physics-based ensemble, although the provided excerpt does not give the size of that increase.