StormNet, OpenSnow’s state-of-the-art severe weather prediction model, scored first place during the spring 2026 NOAA Hazardous Weather Testbed Spring Forecasting Experiment (HWT-SFE) on mid-range severe weather prediction.

StormNet's first place put it ahead of models from the National Center for Atmospheric Research (NCAR), the University of Oklahoma, and Google DeepMind.

The score is based on the ratings of hundreds of expert human forecasters.

StormNet not only came in first place overall, but it also came in first place for each lead day (day 3, day 4, day 5, day 6, and day 7).

A Level Playing Field

StormNet predicts tornadoes, damaging winds, large hail, and lightning from nowcasting timescales (on the order of 10 to 20 minutes) to longer range timescales (out to 14 days).

It is the first model of its kind to separately predict each individual hazard out to 14 days into the future.

Other models, like the models evaluated against StormNet at the HWT-SFE, only predict where severe storms will occur (any hazard instead of specific hazards).

To level the playing field, we combined all of the individual hazard forecasts into a ‘general severe forecast’ for the evaluation.

For more information about StormNet’s development, see our Frequently Asked Questions.

What is the HWT-SFE?

The Hazardous Weather Testbed Spring Forecasting Experiment (HWT-SFE) is an annual evaluation of weather prediction models by the National Severe Storms Laboratory (NSSL) and the Storm Prediction Center (SPC), two branches of NOAA.

The HWT was first established in the early 2000s and has been pivotal in evaluating new models used by NOAA, such as the SSEO, HREF, NSSL-WRF, HRRR, RRFS, REFS, and more.

The experiment runs during the peak severe storm season: late April through May.

Hundreds of meteorologists from across the country converge on Norman, OK (mostly in person, but some virtually as well) to play a role in this experiment. Most of the meteorologist participants are associated with the NWS or universities, and several are from the private sector. Evaluations are performed on physics-based weather models (like the HRRR and RRFS), ensemble methods (like the SSEOv2 and the HREF), AI models (like StormNet).

The participants evaluate models on a subjective 1-10 scale, based on how useful the given model is for a forecast in hindsight.

The evaluation process is blind, so participants are not aware of what model they are evaluating until after the scores are submitted.

The participants see a model’s forecast next to observations (or storm reports) and rate the performance.

StormNet in HWT-SFE

The HWT-SFE evaluated StormNet on mid-range severe storm prediction: probabilistic forecasts for severe thunderstorms with lead times of 3 to 7 days.

The other models in this evaluation were:

• Two versions of GEFS-MLP, originally developed by researchers at Colorado State University and the University of Oklahoma
• AIWP-mean-IFS, a blend of post-processed AI-NWP emulators, Pangu-Weather, and FengWu, initialized with ECMWF’s IFS and developed by NCAR
• AIWP-median-IFS, a blend of post-processed AI-NWP emulators, GraphCast, Pangu-Weather, and FourCastNet, developed by University of Oklahoma researchers
• SUPER-EC, a deep-learning system trained on ERA5, using ECMWF IFS forecasts, and developed by SPC meteorologists
• WeatherNext2, a state-of-the-art AI-NWP emulator developed by Google DeepMind, post-processed into severe storm probabilities

Of all of these models, StormNet scored first place overall, first place each week of the experiment, and first place for each lead time. The facilitators of the experiment were impressed with StormNet’s strong performance.

Real-time StormNet forecasts are available in the StormNet app and at StormNet.ai

Contacts

Andrew Brady, lead StormNet developer: [email protected]

Israel Jirak, NOAA Storm Prediction Center: [email protected]

Matthew Cappucci, a private sector meteorologist who is NOT affiliated with StormNet, finds the model to be the most skillful in the early identification of severe weather setups.