Earth-2 - NVIDIA's open-source family of AI weather forecasting models

What is Earth-2?

NVIDIA Earth-2 is the world's first fully open-source AI weather forecasting model family launched by NVIDIA. Based on three core architectures—Atlas, StormScope, and HealDA—it achieves 15-day global medium-range forecasts, 0-6 hour kilometer-level nowcasts, and second-level initial condition generation, respectively. Compared to traditional numerical weather prediction that relies on supercomputing to solve physical equations, Earth-2 adopts an end-to-end AI pipeline, increasing speed by a thousand times and drastically reducing energy consumption. This enables meteorological agencies, enterprises, and developers around the world to build independent and controllable high-precision forecasting systems based on open-source models, marking the official entry of meteorological AI into the era of industrial deployment.

Earth-2's main functions

• Earth-2 Medium Range Weather ForecastBased on the Atlas architecture, it provides global weather forecasts for up to 15 days and can predict more than 70 meteorological variables such as temperature, air pressure, wind speed, and humidity. In key industry benchmark tests, its accuracy surpasses leading open models such as Google DeepMind's GenCast.
• Earth-2 NowcastingBased on the StormScope architecture, it uses generative AI technology to directly process satellite and radar data to predict short-term extreme weather events from 0 to 6 hours in advance, and can generate local storm forecasts with kilometer-level resolution within minutes.
• Earth-2 Global Data AssimilationBased on the HealDA architecture, it solves the most time-consuming problem of generating initial conditions in weather forecasting. What traditional supercomputers take hours to complete can be done in seconds on a GPU. Combined with intermediate models, it can build the most powerful forecasting pipeline currently available that is fully driven by AI.
• AI downscaling (Earth-2 CorrDiff)The generative AI downscaling method is achieved by using a conditional diffusion model, which is 500 times faster and 10,000 times more energy efficient than traditional methods. It supports refining coarse-grained forecasts to the street level and generates more datasets to better assess the probability distribution of weather events.
• Visualization servicesBased on the NVIDIA Omniverse and OpenUSD 3D framework, it supports the aggregation and visualization of global-scale climate simulations and geospatial datasets, enabling high-resolution interactive exploration in the cloud.

Earth-2's technical principles

• Atlas architectureThe Atlas architecture is the core of the Earth-2 medium-term forecasting model. It uses deep learning networks to process global-scale spatiotemporal series data and replaces the complex physical equation solving process in traditional numerical weather prediction with an end-to-end AI pipeline. It can significantly improve forecasting efficiency while maintaining high accuracy, and achieve global forecasts of more than 70 meteorological variables for up to 15 days.
• StormScope architecture The StormScope architecture is the technological foundation of the Earth-2 near-term prediction model. It uses generative AI technology to learn and infer directly from satellite and radar imagery, without relying on traditional physical equations to simulate atmospheric motion. This allows the model to capture the dynamic evolution of storm systems more quickly and accurately, making it the first AI architecture to surpass traditional physical methods in storm dynamics simulation.
• HealDA architectureThe HealDA architecture is the core innovation of the Earth-2 global data assimilation model. It completely reconstructs the data assimilation process through the parallel computing capabilities of GPUs, compressing the initial atmospheric condition generation task, which traditionally takes several hours to complete on a supercomputer, into a task that takes only seconds. This enables efficient coupling between observational data and prediction models, laying the foundation for building a fully AI-driven end-to-end forecast pipeline.
• CorrDiff technologyCorrDiff technology achieves generative downscaling based on conditional diffusion models. It uses the super-resolution capabilities of generative AI to refine low-resolution atmospheric forecast data to a high-resolution local scale, achieving a 500-fold speed improvement and a 10,000-fold energy efficiency improvement while maintaining physical consistency, making street-level accurate forecasts possible.

Earth-2 project address

• Project official websitehttps://www.nvidia.com/en-us/high-performance-computing/earth-2/
• GitHub repositoryhttps://github.com/NVIDIA/earth2studio
• HuggingFace model libraryhttps://huggingface.co/collections/nvidia/earth-2

Application scenarios of Earth-2

• Weather forecast serviceMeteorological agencies can use local data to fine-tune the Earth-2 open-source model to build a fully independent and controllable national weather forecasting system that conforms to the country's geographical characteristics.
• Extreme Weather WarningBy using models such as FourCastNet, we can predict extreme weather events such as typhoons, rainstorms, and heat waves three weeks in advance, generating 1,000 ensemble forecast members to buy valuable response time for disaster prevention and mitigation.
• Clean Energy ManagementAccurately predict changes in wind farms and solar irradiance, optimize power output forecasting for wind farms and photovoltaic power plants, and enhance the capacity for renewable energy consumption.
• Precision planting in agricultureIt guides farmers to accurately schedule irrigation, fertilization, sowing, and pest and disease control, thereby reducing losses from meteorological disasters and improving crop yield and quality.
• Aviation and shipping safetyReal-time monitoring of flight route weather conditions allows for advance avoidance of dangerous weather such as clear-sky turbulence and thunderstorms, optimizing flight paths and takeoff and landing times to improve on-time performance and flight safety.