NVIDIA has introduced CorrDiff, a generative AI weather model designed to enhance high-resolution weather forecasting. CorrDiff, part of the NVIDIA Earth-2 platform, enables kilometer-scale predictions for wind, temperature, and precipitation, improving the accuracy of extreme weather forecasts. Featured in Communications Earth and Environment, the model refines atmospheric data from a 25-kilometer scale down to 2 kilometers, using AI techniques similar to those found in text-to-image generation. Weather agencies and researchers worldwide are adopting CorrDiff to improve forecasting precision, support disaster planning, and enhance climate resilience efforts. The model, optimized for Taiwan’s weather data in collaboration with the Central Weather Administration, has now been expanded to cover the continental United States.
NVIDIA’s release of CorrDiff as a microservice offers a significant improvement in speed and energy efficiency, running 500 times faster and using 10,000 times less energy than traditional numerical weather prediction methods. The model has been trained on real-world extreme weather events, including hurricanes and tornadoes, to refine forecasting accuracy. Taiwan’s National Science and Technology Center for Disaster Reduction has already integrated CorrDiff into its monitoring systems to improve disaster alerts. NVIDIA continues to enhance the model, exploring applications including urban wind pattern prediction and regional energy forecasting.
