Postprocessing Global AI Weather Prediction Models for Severe Weather Forecasting

Data driven global AI models have garnered significant attention of late owing to their competitiveness with state-of-the-art dynamics-based numerical weather prediction (NWP) systems. After training, the AI models can simulate synoptic-scale patterns with relative ease and speed, and recent quantitative evaluations have suggested they can simulate large-scale temperature, pressure, and winds with the same or better accuracy than NWP models. However, these AI models suffer from some of the same deficiencies as NWP models for applications in high-impact weather domains. Namely, their rather coarse resolution doesn’t lend itself well to explicit predictions of convection hazards. To get around this issue with traditional NWP models, postprocessing methods have been used to generate explicit forecasts of hazards. As an example, the Global Ensemble Forecast System Machine Learning Probabilities (GEFS-MLP) forecast system leverages random forests (RFs) and inputs from a global NWP ensemble to generate daily probabilistic forecasts at lead times of 1-8 days for excessive rainfall, tornado, severe hail, and severe wind hazards. Output forecasts mimic operational outlooks and are now operational in national forecast centers. In a similar way, output fields from the data-driven AI models can be used to generate hazardous weather outlooks. In this work, we apply the previously detailed GEFS-MLP framework to global AI model inputs, to generate daily probabilities of tornadoes, hail, and wind out to 8 days. We explore inputs from PanguWeather, GraphCast, and FourCastNet to drive separate severe weather predictions, and consider combining AI outputs as a 3-member ensemble to drive RF training and forecasts. The operations-like outlooks are compared to similar operational GEFS-MLP products to examine how data driven AI models can be used for small-scale hazardous weather prediction. Additionally, we explore generating ensembles of MLP forecasts by applying trained RFs to individual GEFS ensemble members to understand the value added by generating probabilistic forecasts. Further, explainable AI techniques are leveraged to decipher how MLP systems respond to inputs from different forecast models and whether global AI systems could be used in an ingredients-based forecasting paradigm.