StormCast: kilometer-scale weather with generative deep learning

NERSC, in a research collaboration led by NVIDIA, has recently published a preprint describing a new generative AI weather model called StormCast. StormCast is the first AI-driven weather model capable of emulating the atmosphere at mesoscale (~kilometer-scale) resolution, enabling the ability to represent convective systems and the evolution of extreme weather phenomena associated with them. These include events like thunderstorms, tornadoes, derechos, and extreme precipitation, which are of particular interest over the United States and have substantial and sometimes catastrophic impacts on livelihood, property, and infrastructure. Forecasting these phenomena is difficult because it requires modeling the atmosphere at very high-resolution, which forces tradeoffs on ensemble size or the number of forecasts that can be made at a given time. Further complicating things, most of the available archived observational or ‘analysis’ data is only available on frequencies of 1 hour or greater, while the dynamics of convective systems happen on sub-hourly timescales, which adds uncertainty to the forecasting problem for any AI-based weather model.

StormCast addresses these issues by posing the high-resolution weather forecasting problem as a generative modeling task. Rather than just making a single forecast from a given initial condition, StormCast can generate an ensemble of forecasts, where each ensemble member represents a different possible trajectory of the atmospheric state (and corresponding development, dissipation, or propagation of storms). The model is a generative diffusion model, utilizing some of the same deep learning methods as popular text-to-image generators like DALL-E or Imagen, that has been adapted specifically to work with global and regional weather data for a large number of atmospheric variables (winds, pressure, temperature, humidity, precipitation, etc).

In terms of critical forecast outputs like radar reflectivity, which is heavily used to assess the development of storms and extreme weather, StormCast shows forecast skill competitive with and sometimes exceeding state-of-the-art convective simulators like the High Resolution Rapid Refresh (HRRR) model from NOAA. Furthermore, since StormCast predicts much more than just radar reflectivity, the full atmospheric state can be analyzed, showing that StormCast produces vertical structure at the kilometer-scale that is consistent with the underlying physics for convective motions. This boosts the utility and interpretability of the model for operational or real-time forecast assessments. In general, these results show lots of promise for deep learning applied to regional and extreme weather modeling, and add to the evidence that deep learning and AI will be a transformative tool for weather and climate researchers.

To learn more about StormCast, you can read the preprint, or some of the news coverage and press releases on it, including a technical blog from NVIDIA and coverage from the SF Chronicle and HPCWire.