Extreme weather events pose big concerns for California, from torrential storms to wildfire-fanning offshore winds. Complex, rapidly changing conditions are hard for traditional weather models to calculate, making for challenging forecasts on the ground. A new artificial intelligence weather model from researchers at Nvidia, Lawrence Berkeley National Laboratory and the University of Washington could change that.
“In California, where fire weather is a significant concern, this probabilistic model has a lot of promise,” said Peter Harrington, a machine learning engineer at Lawrence Berkeley National Laboratory.
Harrington and other researchers unveiled StormCast in a preprint study last week, marking one of the first times that generative AI has been able to forecast weather behavior at the level of convective activity within storms. The new approach captures details that, in the future, could help California meteorologists more accurately forecast dangerous fire weather and devastating floods.
The work builds on recent progress with less-detailed AI weather models, which have been “game-changing,” said author Mike Pritchard, who is also the director of climate simulation research for Nvidia.
Traditional weather models rely on expensive supercomputers to crank through complex physics-based equations. AI models, including StormCast, require far less computing power because they instead recognize patterns from past observational data to shortcut the model-making process.
Major weather agencies across the globe have already begun incorporating AI models into their products.
StormCast was trained using several years’ worth of data from the National Oceanic and Atmospheric Administration’s high-resolution rapid refresh (HRRR) model. The HRRR model gives meteorologists key details for tracking the evolution of quickly developing weather phenomena like thunderstorms and how winds will direct wildfire smoke plumes.
The new Nvidia model also uses generative AI to produce fine-scale details of cloud and convection, with even some slight improvements compared with the HRRR model.
The boost may have been because StormCast features an ensemble of five model runs, made possible due to the reduced computational cost with AI models. Ensembles enable more accurate and useful forecasts than models that rely on single solutions, like the HRRR model.
In addition to improving forecasts for fire weather, StormCast could refine predictions for atmospheric river activity, the scientists say. Atmospheric rivers are plumes of tropical moisture that can juice up storms, resulting in downpours that produce devastating floods.
The preprint, which hasn't yet been peer-reviewed, acknowledges limitations. For example, StormCast was designed and tested with weather specifically over the Central United States. Still, Harrington said, “the overall approach we’ve used is totally applicable to all sorts of high-resolution regional weather phenomena.”
StormCast is also a proof of concept, Pritchard said, adding that “these predictions need to be evaluated by card-carrying meteorologists and atmospheric scientists at a scale beyond any given company.”
But StormCast has the potential to revolutionize high-resolution forecasts the way that AI weather models have already improved less-detailed forecasts, Pritchard said: “The same disruption might be imminent.”
(c)2024 the San Francisco Chronicle. Distributed by Tribune Content Agency LLC.
