Nvidia Corp. today detailed an artificial intelligence model called StormCast that could help researchers predict the weather more accurately.

StormCast is an upgraded version of an earlier atmospheric forecasting model dubbed CorrDiff. Nvidia ships the latter algorithm with Earth-2, a software suite it provides to meteorologists who use its chips to support their research. The bundle includes weather forecasting algorithms, tools for managing atmospheric data and related components.

CorrDiff, the model on which StormCast is based, functions as the meteorological equivalent of a zoom-in tool. One way researchers can use it is by uploading a dataset that describes a weather event with a resolution of 25 kilometers. With such a resolution, atmospheric phenomena that have a reach of under 25 kilometers are not visible. CorrDiff can take the raw data and sharpen the resolution by a factor of 12.5 to two kilometers.

StormCast enhances the model’s core feature set by adding a so-called autoregressive capability. The upgrade allows the AI to not only study past weather events but also predict future developments. StormCast generates projections by studying historic atmospheric information, including two and a half years’ worth of climate measurements from the central U.S. that Nvidia included in the model’s training dataset. 

StormCast’s autoregression capability enables researchers to forecast the weather up to six hours into the future. “StormCast enables this at a 3-kilometer, hourly scale,” Mike Pritchard, the director of climate simulation research at Nvidia, detailed in a blog post.  

StormCast is designed to predict so-called mesoscale weather events. Those are atmospheric phenomena with horizontal dimensions that range from five kilometers to several hundred. The category encompasses, among others, flash floods and derechos, long-lasting storms capable of causing extensive wind damage. Regular storms don’t qualify as mesoscale events because they affect a significantly smaller area.

Meteorologists usually forecast the weather using algorithms called convection-allowing models, or CAMs. Such algorithms often run on supercomputers and take upwards of thousands of atmospheric parameters into account to generate predictions. Nvidia says that StormCast has demonstrated the ability to outperform CAM software in some cases. 

“Despite being in its infancy, the model — when applied with precipitation radars — already offers forecasts with lead times of up to six hours that are up to 10% more accurate than the U.S. National Oceanic and Atmospheric Administration (NOAA)’s state-of-the-art 3-kilometer operational CAM,” Pritchard detailed.

Other companies are also researching ways of using AI to improve weather forecasting. Last November, Google LLC detailed GraphCast, an internally developed neural network that can predict atmospheric events faster than traditional algorithms. It has demonstrated the ability to forecast the weather up to 10 days in advance with detailed temperature and wind speed estimates. 

Photo: Unsplash