REV-TOR is being trained on over 40,000 tornadoes (as well as hail & wind events) going back to 1979. All told, over 577,000 severe weather events are being analyzed. Because of the extensive size of the dataset, time it takes to obtain and prepare the data, and the time it takes to train the model, it is being done in chunks and is an ongoing process. Each day the model is getting smarter as new data is ingested. The model is not only learning if a tornado is or is not likely, but also the intensity (F0-F5). The same goes for our concurrent hail and wind variations, providing size of hail and wind speed probabilities.
The atmosphere prior to each event is simulated using (eventually) 36 years of NCEP-NCAR NARR Reanalysis data to model the atmosphere prior to touchdown. We also input more non-event cases than actual tornado cases so the model can better understand the difference between a tornadic atmosphere and a non-tornadic atmosphere.
For each event, over 500 weather variables are analyzed and prioritized within the model, learning of any correlations of that parameter to tornadogenesis. After the AI model 'learns' what variables are and are not important, it will be tweaked by removing variables that may have "dirtied up the water". This will be done by removing variables that are statistically insignificant. Because we can run an analysis that effectively tells us how good each change we make is, we can decide to hone in on the most important variables to maximize model accuracy.
Initially we will be running the early versions of REV-TOR on top of 12z HRRR forecast data for day 1. As we decide which versions of the model are most accurate and settle on a best solution, we will extend these forecasts out to 16 days. We will use our proprietary blend of forecast guidance that uses the best forecast deterministic and ensemble models in the world. This includes the European ECMWF's IFS, AIFS & EPS; Britain's MetOffice UKMET & MOGREPS, the United State's NCEP GFS, GEFS, HRRR, NAM, NBM; Germany's DWD ICON, and Environment Canada's GEM & GEPS to name a few. Our blending method factors in skill of the model, spatial resolution, and leans on each model's strengths in the temporal dimension. This base forecast data is essential to simulating an accurate environment that could potentially develop a tornado.
Not every AI model is the same. How you train the model, the data used, variables asked of the model to utilize, and type & configuration of AI all can make dramatic differences. This is our version. The exact ML configuration we will keep to ourselves. Grandma always said to not give away all your secrets.
This is experimental, and these forecasts are not intended to be used to protect life or property and should only be interpreted by a trained meteorologist who understand the biases and caveats of computer model guidance.
We are currently providing this data free to use. However, donations are appreciated to help us maintain servers and continue doing further development.