Exclusive: New Nvidia model could bolster severe weather forecasts
Add Axios as your preferred source to
see more of our stories on Google.
Illustration: Maura Losch/Axios
A new generative AI weather model from Nvidia Research significantly advances the accuracy of short-range weather forecasting, helping to predict hazards from flash floods to tornado outbreaks, the company first told Axios.
Why it matters: This is the first AI model to demonstrate improved skill at simulating extreme weather events down to the kilometer scale, Nvidia claims.
- If these findings are replicated by other researchers, they could usher in a new era of even more accurate short-term forecasts.
- That could save lives and money by helping people stay safe and protect their property.
The big picture: Until now, AI weather and climate models from Nvidia, Microsoft, Google and researchers elsewhere had demonstrated advances in using AI and machine learning to produce medium-range, global weather projections that rival or beat conventional, physics-based models run on supercomputers.
- In addition to accurate forecasts, the new model, outlined in a study out today as a preprint, could help scientists take global climate change projections and more accurately apply them to local scales.
- This is called "downscaling," and it too has historically been computationally limited.
What they're saying: "I'm convinced we're at that moment now where AI can compete with physics for storm-scale prediction," study coauthor Mike Pritchard, a climate scientist at Nvidia, told Axios.
Zoom in: The new model, discussed in a paper submitted for peer review, is known as StormCast. The paper is written with the Lawrence Berkeley National Laboratory and University of Washington.
- Nvidia claims the model can operate at the mesoscale level, which means it can resolve small-scale weather systems and includes some of the complicated physics of how thunderstorms form, intensify and dissipate.
- Such "convection-allowing models" are used to predict the evolution of derechos, squall lines and deadly flash flood events, among other applications.
- According to the study, the model already beats the accuracy of the top U.S. short-range weather prediction model, known as high-resolution rapid refresh, or HRRR, when it comes to predictions of how storms will evolve on weather radar.
Yes, but: Pritchard cautioned that physics should not give way entirely to generative AI, but could be used to help test new models' "physical credibility that [enables] us to trust AI predictions and to understand their boundaries."
Zoom out: Weather prediction and climate research are in the midst of a rapid shakeup in how insights can be generated.
- AI models have tremendous compute power and timeliness advantages over traditional computer models.
- Experts agree that weather and climate are headed toward greater use of various forms of AI tools, though most — including Nvidia Research —argue that the proven, older ways of doing things should not be discarded.
The intrigue: While AI models, trained on historical weather and climate data, can be run in minutes on easily purchased computers, traditional weather forecasting models take hours to run on the world's most powerful supercomputers.
- Because the new model produces forecasts so quickly, it is possible for it to be used as part of an ensemble system, in which the model is run many times with slightly tweaked initial data to determine how sensitive the forecast is to shifts in different parameters.
- Ensemble forecasting can provide forecasters with more confidence in their predictions or raise red flags that a storm situation may evolve differently.
Between the lines: Nvidia, which is the leading maker of AI-enabling chip technology, has invested in pushing scientific boundaries in a variety of fields, including weather and climate.
- It has created a digital twin of Earth, which Taiwan plans to use for improving typhoon predictions.
- StormCast produces hourly predictions at small geographic scales, using generative AI trained on data of historical observations and forecast outcomes.
- Nvidia Research claims it is the first AI weather simulation to predict the variables, such as temperature and moisture concentration, that determine atmospheric buoyancy — or the tendency to produce lift.
- This is a critical advance since buoyancy can make or break a group of thunderstorms' ability to produce tornadoes, congeal into even large complexes that yield flooding rainfall, or weaken altogether. It is also a key factor for predicting heavy bands of snow during the winter.
What's next: Nvidia is working with The Weather Company as well as Colorado State University to test the new model, and may make it more widely available.
The bottom line: "An entirely new way to simulate the atmosphere has been born in the last five years," Pritchard said. "We're sort of scratching the tip of the iceberg on these local scales, what AI will be able to do. So I think it's a time to be excited about the potential improvements in local weather simulation."
