NVIDIA Earth-2
Last reviewed
Jun 7, 2026
Sources
17 citations
Review status
Source-backed
Revision
v1 · 1,799 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 7, 2026
Sources
17 citations
Review status
Source-backed
Revision
v1 · 1,799 words
Add missing citations, update stale details, or suggest a clearer explanation.
NVIDIA Earth-2 is a software platform and family of models from NVIDIA for AI-accelerated weather and climate simulation, prediction, and visualization. It combines physics-based numerical simulation, AI emulators trained on observational and reanalysis data, and interactive 3D visualization, with the stated long-term goal of building a kilometer-scale "digital twin" of Earth's atmosphere. NVIDIA first announced the Earth Climate Digital Twin under the Earth-2 name on March 18, 2024, and on January 26, 2026 released the Earth-2 family of open models, which NVIDIA describes as the first fully open, accelerated weather AI software stack spanning the entire forecasting pipeline. Earth-2 is the weather and climate pillar of NVIDIA's broader AI for science push, and it is best understood as an open ecosystem of models, libraries, and reference tools rather than a single forecast product.
Earth-2 brings together GPU acceleration, generative AI, traditional physical simulation, and computer graphics to simulate and visualize weather at fine spatial scales. The original 2024 announcement framed Earth-2 around an Omniverse-based interactive digital twin able to render and simulate atmospheric conditions toward a 2-kilometer resolution, motivated by the rising frequency of extreme-weather disasters. NVIDIA chief executive Jensen Huang described the effort at launch as a way to help society "better prepare for, and inspire us to act to moderate," extreme weather.
Over 2024 and 2025 the platform expanded from a visualization-and-downscaling concept into a layered software stack. The lower layers include PhysicsNeMo (formerly Modulus), a framework for training physics-informed and deep learning weather models, and Earth2Studio, an open-source Python toolkit for assembling inference pipelines that chain forecasting, downscaling, and post-processing models together. On top of these sit the trained models themselves. A distinction worth keeping in mind is that Earth-2 hosts and standardizes both NVIDIA's own models and third-party models, so the platform is a hosting and tooling environment as much as a model zoo.
The two longest-running NVIDIA models in Earth-2 are FourCastNet and CorrDiff, which address different parts of the problem.
FourCastNet (Fourier Forecasting Neural Network) is a global, data-driven medium-range forecast model first published in 2022 by researchers from NVIDIA, Lawrence Berkeley National Laboratory, the University of Michigan, and Rice University. It originally used Adaptive Fourier Neural Operators trained on the ECMWF ERA5 reanalysis, producing global forecasts at 0.25-degree resolution. The latest version, FourCastNet 3 (FCN3), was published on July 29, 2025 and moves to a spherical neural operator architecture built on spherical signal-processing primitives, with stochasticity introduced through a latent diffusion process on the sphere to generate calibrated ensembles. Per NVIDIA's technical reporting, a single 60-day FCN3 rollout at 0.25 degrees and 6-hourly resolution runs in under four minutes on one NVIDIA H100 GPU, which NVIDIA cites as roughly an 8x speedup over Google DeepMind's GenCast and about 60x faster than ECMWF's IFS-ENS ensemble.
CorrDiff is a generative diffusion model for downscaling, also called super-resolution. Introduced at GTC 2024 and described in a peer-reviewed Communications Earth and Environment paper, CorrDiff uses a two-step approach in which a deterministic model predicts the mean field and a diffusion model adds fine-scale correction, taking coarse roughly 25-kilometer input down to about 2-kilometer regional detail while also performing bias correction and variable synthesis. NVIDIA reports CorrDiff running on the order of 500x to 1,000x faster and thousands of times more energy-efficient than comparable CPU-based numerical approaches; the precise multiplier varies by configuration and source, so these figures are best read as NVIDIA's own benchmark claims rather than independent results. The initial CorrDiff model was tuned on Taiwan weather data with Taiwan's Central Weather Administration.
The platform also includes cBottle (Climate in a Bottle), announced June 10, 2025, which NVIDIA calls the first generative AI foundation model designed to simulate global climate at kilometer scale. cBottle is trained on high-resolution physical climate simulations plus roughly 50 years of observation-constrained atmospheric estimates, and can fill gaps, super-resolve coarse data, and correct biased climate models. DLESyM, a deep-learning Earth-system emulator, is another component referenced in the open stack.
At the American Meteorological Society Annual Meeting in Houston, NVIDIA on January 26, 2026 launched the Earth-2 family of open models, positioning it as an end-to-end, fully AI pipeline from raw observations to local storm prediction. Three new models, each named by its architecture, anchor the release, alongside the existing CorrDiff and FourCastNet 3.
Earth-2 Nowcasting, built on an architecture called StormScope, uses generative AI to turn country-scale inputs into kilometer-resolution, zero- to six-hour predictions of local storms and hazardous weather, generating satellite and radar-style imagery directly from geostationary satellite observations such as GOES, initially over the contiguous United States. NVIDIA states it is the first AI model to outperform traditional physics-based systems on short-term precipitation forecasting by simulating storm dynamics directly.
Earth-2 Medium Range, built on an architecture called Atlas, is a latent diffusion transformer trained on ERA5 that produces global forecasts up to 15 days ahead across more than 70 weather variables, predicting incremental atmospheric changes to preserve fine structure. NVIDIA reports it outperforms GenCast on standard benchmarks; this is a vendor claim pending independent evaluation.
Earth-2 Global Data Assimilation, built on an architecture called HealDA, produces the initial atmospheric conditions that forecasts start from, generating global snapshots of temperature, wind speed, humidity, and pressure in seconds on GPUs rather than hours on supercomputers. NVIDIA says pairing HealDA with Medium Range yields the most skillful fully open, fully AI weather pipeline available, and that this model would arrive later in 2026 while Nowcasting and Medium Range shipped at launch.
| Model / component | Architecture codename | Role | Resolution / range | First released |
|---|---|---|---|---|
| Earth-2 Nowcasting | StormScope | Convective nowcasting from satellite imagery | Kilometer-scale, 0 to 6 hours (CONUS) | January 2026 |
| Earth-2 Medium Range | Atlas | Global medium-range forecasting | 70+ variables, up to 15 days | January 2026 |
| Earth-2 Global Data Assimilation | HealDA | Initial-condition generation | Global, seconds on GPU | Expected later in 2026 |
| FourCastNet 3 | FCN3 (spherical neural operator) | Global ensemble forecasting | 0.25 deg, up to 60 days | July 2025 |
| CorrDiff | Correction diffusion | Downscaling / super-resolution | ~25 km to ~2 km | March 2024 (GTC) |
| cBottle | Climate in a Bottle | Generative climate foundation model | Kilometer-scale global | June 2025 |
The 2026 models ship through Earth2Studio, Hugging Face, and GitHub as open, customizable checkpoints intended for sovereign and on-premises deployment, with accompanying research papers.
NVIDIA has built Earth-2 around named adopters rather than a single flagship customer. In a March 18, 2025 announcement, climate-technology companies and agencies cited as using Earth-2 or CorrDiff included G42, Spire Global, JBA Risk Management, Tomorrow.io, GCL, Esri, Ecopia, OroraTech, and Taiwan's Central Weather Administration. G42, through its Inception unit, built a custom downscaling system on CorrDiff for the United Arab Emirates' National Center of Meteorology, and Spire reported sub-seasonal ensemble forecasting it characterized as roughly 1,000x faster than physics-based models. Taiwan's National Science and Technology Center for Disaster Reduction embedded CorrDiff in its disaster-monitoring site to help forecasters prepare for typhoons, with NVIDIA citing an estimated gigawatt-hour of energy savings.
The 2026 open-model launch named a wider set of users, including the Israel Meteorological Service, The Weather Company, the U.S. National Weather Service, TotalEnergies, Eni, Southwest Power Pool, AXA, and S&P Global Commodity Insights, spanning national agencies, energy firms, and insurers. Amir Givati of the Israel Meteorological Service was quoted saying the Earth-2 models gave a 90 percent reduction in compute time at 2.5-kilometer resolution. On the research side, the Max-Planck-Institute for Meteorology and the Allen Institute for AI (Ai2) have explored cBottle. The Weather Company has separately discussed integrating its meteorological data and tools with NVIDIA Omniverse for digital-twin visualization.
Earth-2 sits at the intersection of two trends: the rapid rise of AI weather models that rival or beat traditional numerical weather prediction at a tiny fraction of the compute, and NVIDIA's strategy of supplying the full hardware-plus-software stack for scientific computing. By open-sourcing models across the whole pipeline, from data assimilation through medium-range forecasting, nowcasting, and downscaling, NVIDIA lowers the barrier for national meteorological agencies, startups, and researchers to run frontier forecasting on their own infrastructure, while keeping that work anchored to NVIDIA GPUs and libraries.
The main caveats are interpretive. Many of the headline accuracy and speed numbers, such as outperforming GenCast or IFS-ENS and the large energy-efficiency multipliers, originate from NVIDIA's own benchmarks and await broad independent confirmation from the meteorological community, which evaluates new models against shared baselines like ECMWF scorecards. The "digital twin" framing is also aspirational: a continuously running, fully coupled, 2-kilometer global twin remains a goal rather than a deployed product, and Earth-2 today is more accurately a toolkit for building regional twins and forecast pipelines. Even with those qualifications, the January 2026 release marked one of the first times a complete, openly licensed AI weather stack became available end to end, which is a meaningful shift for operational forecasting.