Ensemble Forecasting with Global Machine Learning Models

In the past years, global machine learning weather models trained on ERA5 reanalysis have shown remarkable ability to produce weather forecasts akin to state of the art global models (IFS and GFS) but with speedup at the order of 10,000x (see ECMWF blog post). One of the immediate applications of such a speedup is ensemble forecasting, where traditionally computational limitations force a tradeoff between the ensemble size and the model resolution.

In this presentation various techniques for ensemble forecasting, tailored for machine learning weather models, are explored. Initial condition uncertainty is represented with different noise patterns respecting the spherical geometry while model uncertainty is represented with either noise injection between steps or composing an ensemble of different trained models of similar skill. The optimization of the ensemble setting to probabilistic forecasting of specific variables is done by exploring parameters (amplitudes and spatial scales) and by calibrations to a spread/skill ratio of 1. We explore perturbing subset of model variables at a subset of the levels and how this perturbation propagates to other variables with model steps to better understand the response of the model to perturbations.

Results show 1000 member ensemble predictions of weather extremes, such as heat waves, atmospheric rivers and tropical cyclones. Probabilistic metrics like CRPS, spread skill ratio and extreme forecast index are examined and the implications of different perturbation methods on the predicted variables are discussed in light of these results.