NVIDIA releases StormCast, a weather forecasting model, AI makes "chasing winds" faster and more accurate
2024-08-24
한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina
Typhoons, a "familiar face" in nature, often have a huge impact on people's production and life with their amazing destructive power and unpredictable paths. Every year from June to August, it is also the active period of typhoons in the northwest Pacific. According to the forecast of China Weather Network, according to the "schedule" of typhoons in August this year, 3 to 4 typhoons will be generated in the northwest Pacific and South China Sea in August, which is less than the same period of previous years. Among them, 2 to 3 typhoons will land in my country, and the typhoons will mainly move westward and northwestward.
In recent years, the technology and tools for meteorological "wind chasing" have also been continuously evolving and upgrading. The reporter learned that NVIDIA recently released StormCast, a breakthrough generative AI model for simulating high-fidelity atmospheric dynamics. The breakthrough point of this model is that it can make reliable predictions of weather systems before the formation of storms and before cyclones, making typhoon predictions one step earlier - it can now achieve a spatial resolution of 3 kilometers and a temporal resolution of one hour.
Two months ago, NVIDIA founder and CEO Jensen Huang released the CorrDiff model available through Earth-2. According to reports, NVIDIA Earth-2 is a digital twin cloud platform that integrates AI, physical simulation, and computer graphics, and can simulate and visualize weather and climate forecasts on a global scale with unprecedented accuracy and speed. CorrDiff can increase the model resolution from 25 kilometers to 2 kilometers, which is 12.5 times the previous resolution, and the single inference speed is 1,000 times faster than traditional methods, and the energy efficiency is improved by 3,000 times.
StormCast adds hourly autoregressive forecasting to CorrDiff, which means the model can predict future results based on past results.
Forecast up to 6 hours in advance
The physical hazards caused by weather and climate change vary significantly from region to region, but reliable numerical weather forecasts at this scale are computationally expensive because of the high spatial resolution required to simulate basic fluid dynamics at the mesoscale.
Therefore, in the past, meteorological researchers have often had to make trade-offs between resolution, ensemble size, and affordability in regional weather forecast models (usually called convection-allowed models, abbreviated as CAMs).
At lower resolutions, machine learning models trained on global data have been able to effectively simulate numerical weather prediction models to improve early warning systems for severe events. These machine learning models typically have a spatial resolution of around 30 kilometers and a temporal resolution of 6 hours.
With the help of generative diffusion technology, StormCast is able to achieve a spatial resolution of 3 kilometers and a temporal resolution of one hour. When used in conjunction with precipitation radar, the model can already provide forecasts up to 6 hours in advance.
Additionally, StormCast’s output displays physically realistic heat and humidity dynamics and is able to predict more than 100 variables, such as temperature, moisture concentration, wind, and rainfall radar reflectivity values at multiple fine-scale altitude levels. This allows meteorological researchers to confirm the true 3D evolution of storm buoyancy in AI weather simulations for the first time.
More innovations are in the pipeline
Scientists are already exploring how to take advantage of the model. “Producing computationally tractable storm-scale ensemble forecasts is a formidable challenge in numerical weather prediction, given the large impacts of organized thunderstorms and winter precipitation and the difficulty of reliably forecasting them,” said Tom Hamill, head of innovation at The Weather Company. “StormCast is a clear model that can meet this challenge. The Weather Company is excited to work with NVIDIA to develop, evaluate and use these deep learning forecast models in the future.”
“Developing high-resolution weather models requires using AI algorithms to account for convection, which is a daunting challenge,” said Imme Ebert-Uphoff, head of machine learning at Colorado State University’s Cooperative Institute for the Atmosphere. “This new research from NVIDIA explores the potential of using diffusion models like StormCast to achieve this goal, and represents an important step toward developing AI-powered high-resolution weather forecast models for the future.”
These breakthroughs accelerate and visualize physics-accurate climate simulations and create a digital twin of the Earth. They also demonstrate how NVIDIA Earth-2 is ushering in a new and important era of climate research.
It is reported that NVIDIA Research has hundreds of scientists and engineers around the world, focusing on research in areas such as climate AI, computer graphics, computer vision, self-driving cars and robotics.
Author: Zhang Tianchi
Text: Zhang Tianchi Editor: Zhang Yi Editor-in-Chief: Rong Bing
Please indicate the source when reprinting this article.
