Medium-range Real-time Convective Hazard Forecasts

About the AI-NWP-based Hazard Forecasts

• The Day 1-8 AI-NWP-based convective hazard forecasts are based on output from two 52 member ensembles generated with two different AI NWP emulators (PanguWeather and FengWu) and run on NSF NCAR supercomputing resources.
• This 104-member multi-model ensemble is initialized daily at 00 UTC using the 51-member 00 UTC ECMWF ensemble, plus an additional member initialized with the ECMWF HRES.
• Hazard forecasts are then generated for each member with a decoder-only transformer. Each transformer, one for each model, was trained to take Pangu/FenguWu output and generate probabilities for the occurrence of >= 1 convective weather report (tornado, hail, or wind) within 24-hr and 40-km of a point.
• The transformers were trained with 5 years of 10-day deterministic forecasts from 2018-2023 initialized with the 00 UTC ECMWF HRES.
• NEW for 2026: Hazard forecasts are being generated with the Weathernext2 64-member ensemble mean forecast, produced in real-time by Google DeepMind. Post-processing for hazard prediction done at NSF NCAR using the same transformer-based approach described above, training on 2022-2024 WeatherNext2 mean forecasts.

About the NWP-based Hazard Forecasts

• GEFS-ML RF is the operational Colorado State University (CSU) medium-range hazard probability product. It is trained with random forests (RFs) and uses the operational GEFS. GEFS-ML NN is a similarly derived product developed at NCAR, but uses dense neural networks to make predictions.
• MPAS ML-NN is based on a real-time MPAS 8-member, 5.5 day, ensemble system running daily between 23 April - 30 May 2025. It uses a dense neural network to make Day 1 - 5 predictions of convective hazards. The MPAS ensemble forecasts are available here.