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:
- Applying AI Weather Models With NVIDIA Earth-2: This training lab will show participants how to run global AI weather forecasting models.
- Earth to AI: This panel brings together industry leaders to explore how AI and climate science are transforming business strategies for a sustainable future.
- Enhancing Photovoltaic Power Predictions With High-Resolution Weather Forecasting from NVIDIA Earth-2: This session covers a project between NVIDIA, Peking University and power company GCL to use Earth-2 models to predict solar power generation output.
- Global Atmospheric Downscaling by Improving CorrDiff Process: In this poster, South Korean startup NoteSquare describes a project that modified and applied CorrDiff to regional weather data from the Korea Meteorological Administration.
- Transform Natural Catastrophe Risk Simulations With Advanced Computational Tools: Presenters from NVIDIA, Amazon Web Services and multinational insurance corporation AXA will share how AXA uses Earth-2 to simulate extreme weather.