Nvidia Research, the research arm of Nvidia, has announced a new generative artificial intelligence (AI) model called StormCast, designed to emulate high-fidelity atmospheric dynamics.
This development enhances weather prediction at the mesoscale, a scale larger than storms but smaller than cyclones, which is crucial for disaster planning and mitigation.
Nvidia emphasises the growing importance of improving climate research and prediction through advanced technology as extreme weather events become more frequent and severe.
StormCast is part of Nvidia’s broader effort to leverage AI in climate research. The company’s Earth-2 platform, a digital twin cloud system, integrates AI, physical simulations, and computer graphics to simulate and visualise weather and climate predictions on a global scale with high accuracy.
In Taiwan, the National Science and Technology Center for Disaster Reduction uses Nvidia’s AI models to predict typhoon details, highlighting the efficiency and accuracy improvements offered by such technology. StormCast builds on these advancements by adding hourly autoregressive predictions, enabling forecasts based on past data, which could be critical in improving early-warning systems.
Nvidia joins race to develop AI climate models
According to Nvidia, regional climate research often serves as the foundation for global studies. The physical hazards of weather and climate can vary widely across regions, necessitating accurate regional weather prediction, which comes with high computational costs due to the need for detailed spatial resolution.
Convection-allowing models (CAMs) are essential for tracking storm evolution and structure. They help meteorologists monitor how a storm organises when it forms, which can indicate tornado risk. CAMs also assist researchers in assessing potential hazards to infrastructure.
Global climate models can inform CAMs, enabling them to predict flash floods in coastal areas based on changes in atmospheric moisture. Machine learning models trained on global data have become valuable tools for improving early-warning systems for severe weather events.
StormCast, with generative diffusion, advances these capabilities to a 3km, hourly scale. Despite being in its early stages, the model, when combined with precipitation radars, offers six-hour forecasts that are 10% more accurate than the US National Oceanic and Atmospheric Administration’s (NOAA) 3-kilometre operational CAM.
StormCast’s outputs also display realistic heat and moisture dynamics, said Nvidia, predicting over 100 variables, including temperature, moisture concentration, wind, and rainfall radar reflectivity at multiple altitudes. This enables scientists to observe a storm’s evolution in three dimensions.
Google, Microsoft and IBM pursuing similar platforms
Nvidia trained StormCast on three-and-a-half years of NOAA climate data from the central US, utilising the firm’s accelerated computing to speed up calculations.
“StormCast is a notable model that addresses the challenges of producing storm-scale ensemble weather forecasts,” said The Weather Company’s head of innovation, Tom Hamill. The firm, Hamill added, “is excited to collaborate with Nvidia on developing, evaluating, and potentially using these deep learning forecast models.”
Other big tech companies like Google, Microsoft, and IBM are also advancing AI-driven climate research.
Google has developed the NeuralGCM, a hybrid AI model that combines AI with traditional atmospheric physics to enhance weather forecasting precision and efficiency. This model has shown superior performance in predicting long-term climate trends and extreme weather conditions.
Microsoft has developed medium-range forecast models by integrating advanced AI algorithms with traditional meteorological data. IBM, in partnership with NASA, is developing the Prithvi-weather-climate foundation model. The IBM model uses AI to analyse extensive climate data to improve the accuracy of weather forecasts and other climate-related applications.
Read more: How quantum deals with heavy weather
