AI weather and climate prediction is more accessible than ever for scientists, startups, developers, enterprises and government agencies worldwide.
At the American Meteorological Society’s Annual Meeting, Nvidia this week unveiled a new Nvidia Earth-2 family of open models, libraries and frameworks for weather and climate AI, offering the world’s first fully open, accelerated weather AI software stack.
The Nvidia Earth-2 Nowcasting model uses generative AI trained on satellite and radar data to predict the evolution of realistic cloud and rainfall systems, learning to forecast how storms develop and organize.
These open technologies — including pretrained models, frameworks, customization recipes and inference libraries — accelerate all forecasting stages, from processing initial observation data to generating 15-day global forecasts or local storm forecasts.
Historically, weather forecasting has relied on powerful supercomputers running physics-based models. AI-powered weather forecasting saves significant computational time and costs, allowing more nations, weather enterprises and businesses to run application-specific forecasting systems.
Making production-ready weather AI fully accessible for organizations to run, fine-tune and deploy on their own infrastructure, Nvidia Earth-2 is the first open, accelerated set of models and tools that enables developers to bring disparate weather and climate AI capabilities together.
Ensembles of forecasts of different weather variables — total column water vapour, wind speed and specific humidity — using Nvidia Earth-2 Medium Range, compared against ERA5 reanalysis data.
It’s pioneering work to speed weather prediction, enhance forecasting accuracy, foster collaboration and advance scientists’ overall understanding of the planet’s atmospheric conditions.
Developers across industries are tapping into Earth-2 to predict weather and harness actionable insights. This includes AI weather tool provider Brightband; weather forecasters the Israel Meteorological Service, Taiwan’s Central Weather Administration, The Weather Company and the US National Weather Service (NWS); energy forecasting and grid operations companies TotalEnergies, Eni, GCL and Southwest Powerpool in collaboration with Hitachi; energy trading solutions providers Jua and Metdesk; and financial risk and intelligence firms AXA, JBA Risk Management (The Flood People) and S&P Global Energy.
New Open Models Join Earth-2 Family
The new Earth-2 open weather models are:
- Earth-2 Medium Range, powered by a new model architecture called Atlas, which enables high-accuracy weather prediction for medium-range forecasts — or forecasts of up to 15 days in advance — across 70+ weather variables including temperature, pressure, wind and humidity. On standard benchmarks, it outperforms leading open models on the most common forecasting variables measured by the industry. Read the research paper.
- Earth-2 Nowcasting, powered by a new model architecture called StormScope, which uses generative AI to make country-scale forecasts into kilometer‑resolution, zero- to six-hour predictions of local storms and hazardous weather in just minutes. Earth-2 Nowcasting is the first to outperform traditional, physics-based weather-prediction models on short-term precipitation forecasting by simulating storm dynamics directly. It harnesses AI to directly predict satellite and radar imagery. Read the research paper.
- Earth-2 Global Data Assimilation, powered by a new model architecture called HealDA, which produces initial conditions for weather prediction — snapshots of the current atmosphere, including the temperature, wind speed, humidity and air pressure, at thousands of locations around the globe. Earth-2 Global Data Assimilation can generate initial conditions in seconds on GPUs instead of hours on supercomputers. When coupled with Earth-2 Medium Range, this results in the most skillful forecasting predictions produced by an open, entirely AI pipeline. Read the research paper.
These join existing open weather models in the Nvidia Earth-2 stack:
- Earth-2 CorrDiff, which uses a generative AI architecture called CorrDiff to downscale coarse-resolution, continental-scale predictions to high-resolution, regional-scale weather fields — providing the fine-grain resolution needed for local forecasting up to 500x faster than traditional methods.
- Earth-2 FourCastNet3, which delivers high forecasting accuracy for various weather variables, such as wind, temperature and humidity, surpassing leading conventional ensemble models and rivaling top diffusion-based methods while producing forecasts up to 60x faster than these approaches.
Earth-2 also integrates open models from the European Centre for Medium-Range Weather Forecasts (ECMWF), Microsoft, Google and others. In addition, Earth-2 models can be trained and fine-tuned using Nvidia PhysicsNeMo, an open-source Python framework for developing AI-physics models at scale.
Nvidia Earth-2 Global Data Assimilation shows the complex patterns of Earth observation data from satellites, weather balloons and weather stations, which the AI model transforms into smooth, continuous estimates of the atmospheric state from which forecasts can be launched.