A Hybrid Machine Learning system for high resolution large ensemble coupled forecasts

The objectives of achieving (1) high-resolution global forecasts, (2) the potential for very-large ensembles of O(1,000) to O(10,000), and (3) the potential need to use coupled Earth system models, are addressed by hybridizing AI/ML models with conventional NWP processes. Hybridization is carried out using two separate but compatible approaches: first, a hybrid mixed-model ensemble is developed, with one or more NWP ensemble members and many AI/ML members, and second, the atmosphere component of a conventional coupled atmosphere-wave-ocean NWP model is replaced with an AI/ML atmosphere model that passes flux information to the wave and ocean components at the model time step. For validation, all forecasts are tested against independent satellite and in situ observations to ensure that all forecasts are being compared against an independent 'ground truth' that is not derived from the same NWP system used to train the AI/ML weather models (e.g. the ERA5 reanalysis or ECMWF HRES analysis). The outcome is a system that can produce skillful high-resolution forecasts using large ensembles, validated against observations, with reduced cost compared to a conventional ensemble forecasting system.