Hybrid weather prediction using spectral nudging toward machine-learning forecasts

Abstract

A hybrid approach to numerical weather prediction is investigated, in which the unperturbed physics-based ECMWF Integrated Forecasting System (IFS) is spectrally nudged toward forecasts from a machine-learned weather forecast model, trained to forecast on model levels. Nudging is applied only to the large scales of virtual temperature and vorticity, with the objective of improving large-scale forecast skill while preserving the dynamical and physical behaviour of the underlying physics-based model at smaller scales. Consistent with previous studies, spectral nudging substantially improves large-scale forecast skill relative to the free-running IFS, with gains of up to 1.5 days in the tropics and 12-18 hours in the extra-tropics, and a reduced frequency of forecast busts. These improvements are achieved while preserving forecast variability. The representation of extreme near-surface weather is maintained or improved. Tropical cyclone track forecasts benefit from improved large-scale steering flow, while storm intensity remains comparable to that of the physics-based model and more physically consistent than in pure machine-learned weather forecast models. These results confirm that scale-selective spectral nudging provides a practical pathway for combining machine-learning and physics-based forecasting systems.