Nvidia researchers have developed an AI model called "StormCast" that enables thunderstorm forecasts on a scale of a few kilometers.
Until now, it has been challenging for AI models to capture the complex dynamics of the atmosphere at this fine scale. The researchers combined two innovations for StormCast: First, they rely on a generative model that can simulate many possible developments. Second, StormCast predicts a dense atmospheric state with dozens of vertical layers.
StormCast mimics the high-resolution model "High-Resolution Rapid Refresh" (HRRR) currently used by the American weather service NOAA. It forecasts 99 state variables on a scale of 3 kilometers in hourly time steps, with a particular focus on the lowest layers of the atmosphere.
Nvidia's StormCast is on a different level
In tests, StormCast showed a similar forecast quality to the HRRR model. The probabilities for light, moderate, and heavy rain matched well up to 6 hours in advance. StormCast also realistically reproduced the development of thunderstorm cells, updrafts, downdrafts, and cold air flows under thunderstorms.
One advantage of the AI approach is that ensembles, i.e., groups of slightly varied forecasts, can be easily created. With just five ensemble members, StormCast outperformed the single HRRR run. Such ensembles are very computationally intensive with classical weather models.
Some challenges remain: Among other things, future models should learn on more training data and larger areas. The calibration of the ensembles can also be improved.
Nevertheless, the team is convinced that the results pave the way for a new generation of high-resolution, AI-supported weather models. Such models could help meteorologists predict dangerous thunderstorms even more accurately and quickly, thus avoiding damage and fatalities. The models are also promising for local climate forecasts.
