Nvidia's CorrDiff: Revolutionizing Local Weather Forecasting with AI

Nvidia's CorrDiff uses AI to generate higher resolution local weather forecasting

A team of engineers and weather specialists at Nvidia, working with a colleague from Taiwan's Central Weather Administration, has developed a new AI app aimed specifically at generating higher resolution local weather forecasting. In their paper published in Communications Earth & Environment, the group describes their two-step approach to developing a better local weather forecasting system. Over the past several decades, weather forecasting has improved dramatically, at least in developed countries. Advances in software, hardware and weather modeling have led to the creation of numerical-based modeling systems running on massive supercomputers that produce extremely accurate weather forecasts -- in a macro sense. Local weather prediction is still waiting for improvement, especially in places distant from large metropolitan areas, due to the huge cost that would be involved in using supercomputers to make such predictions. The team at Nvidia has developed a weather forecasting system they call Corrective Diffusion (CorrDiff) that combines the best features of the massive computer models with artificial intelligence. It works by downscaling global weather predictions to a more local level, and then improving their resolution -- and does so at far lower cost than traditional systems. The two-step approach taken by the team involves the use of a deterministic AI mode that produces output based on a given input -- no randomness is involved. The predictions it makes are based on weather behavior patterns that have been learned over time. The second step involves fine-tuning the output from the first step using a generative diffusion model. By using the same basic repetitive techniques used by chatbots to learn how to answer queries more intelligently, the system produces higher and higher resolution results. The result is a system that brings macro-scale accuracy to the local or regional level. The system was tested against several conventional models and was found to provide similar results -- the difference was in the much lower cost associated with such results and the speed with which it delivered them. The team at Nvidia suggests their system can bring accurate forecasting to the local level for people around the world, helping to better predict dangerous weather, and possibly, save lives. In their blog, representatives of the company report that the system is already being used by several meteorological agencies and companies around the world.

NVIDIA Earth-2 Features First Gen AI to Power Weather Super-Resolution for Continental US

Your browser doesn't support HTML5 video. Here is a link to the video instead. To better prepare communities for extreme weather, forecasters first need to see exactly where it'll land. That's why weather agencies and climate scientists around the world are harnessing NVIDIA CorrDiff, a generative AI weather model that enables kilometer-scale forecasts of wind, temperature, and precipitation type and amount. It's part of the NVIDIA Earth-2 platform for simulating weather and climate conditions. The paper behind CorrDiff was featured today in Communications Earth and Environment, part of the Nature portfolio of scientific journals. Available as an easy-to-deploy NVIDIA NIM microservice, the model is already being used by weather technology companies, researchers and government agencies to enhance their forecasts. With the rising frequency of extreme weather events, fast, high-resolution predictions of weather phenomena could help mitigate risks to people, communities and economies by supporting risk assessment, evacuation planning, disaster management and the development of climate-resilient infrastructure. Weather agencies and startups across the globe are adopting CorrDiff and other Earth-2 tools to improve the resolution and precision of forecasts for extreme weather phenomena, renewable energy management and agricultural planning. High-Fidelity Forecasts on the Horizon CorrDiff uses generative AI to sharpen the precision of coarse-resolution weather models -- resolving atmospheric data from 25-kilometer scale down to 2 kilometers using diffusion modeling, the same kind of AI model architecture that powers today's text-to-image generation services. In addition to boosting image resolution, CorrDiff can also predict related variables that weren't present in the input data -- such as radar reflectivity, which is used as an indicator of rain location and intensity. CorrDiff was trained on the Weather Research and Forecasting model's numerical simulations to generate weather patterns at 12x higher resolution. The initial CorrDiff model, announced at NVIDIA GTC 2024 and described in the Communications Earth and Environment paper, was optimized on Taiwan weather data in collaboration with its Central Weather Administration. NVIDIA researchers and engineers next worked to efficiently scale the model to cover a larger section of the globe. The version released as an NVIDIA NIM microservice at Supercomputing 2024 covers the continental United States -- trained on U.S. weather data, with sample datasets of real-world natural disasters including hurricanes, floods, winter storms, tornados and cold waves. The optimized CorrDiff NIM microservice for U.S. data is 500x faster and 10,000x more energy-efficient than traditional high-resolution numerical weather prediction using CPUs. The research team behind CorrDiff continues to advance the model's capabilities, and has released additional generative AI diffusion models showing how the model could be enhanced to more robustly resolve small-scale details in different environments -- and better capture rare or extreme weather events. CorrDiff could also help with downwash prediction -- when strong winds funnel down to street level, damaging buildings and affecting pedestrians -- in urban areas. Weather Agencies Put CorrDiff on the Map Meteorological agencies and companies around the globe are tapping CorrDiff to accelerate predictions with applications in regional forecasting, renewable energy and disaster management. Taiwan's National Science and Technology Center for Disaster Reduction, for instance, has deployed the CorrDiff to support disaster alerts in the region, enabling an estimated gigawatt-hour of energy savings due to the energy efficiency of CorrDiff running on the NVIDIA AI platform. CorrDiff predictions are embedded in the center's disaster monitoring site, helping Taiwan forecasters better prepare for typhoons. Discover Earth-2 at NVIDIA GTC Learn more about AI applications using Earth-2 at NVIDIA GTC, the global AI conference taking place March 17-21 in San Jose, California. Relevant sessions include:

Nvidia unveils generative AI model for local weather forecasts

Why it matters: The new model allows for more efficient and rapid simulations of upcoming weather at ultra-high resolution. This could help significantly improve forecasts, especially those in the short- to medium-range. Driving the news: A study on the model, known as "NVIDIA CoreDiff," and its capabilities was published Monday in the journal Communications Earth & Environment. The big picture: The model is one in a series of AI forecasting advances during the past few years that are poised to remake how weather and climate hazards are predicted. AI forecasting methods train models based on how the atmosphere behaved when certain data parameters were present. The intrigue: The new model is a generative AI model that allows for local-level forecasts of wind, temperature and precipitation type and amounts, the company said in a statement. What they're saying: Mike Pritchard, Nvidia's head of climate simulation research, described the model as "a significant development." Zoom out: Forecasters at the National Weather Service, in the media and elsewhere increasingly are using AI models to augment traditional, physics-based forecast tools. Yes, but: Such expectations could change as AI-based forecasting rapidly evolves and could yield some surprise breakthroughs.