Quick start

Running AI weather prediction can be done with just a few lines of code.

• For detailed installation steps, including model-specific installations, see the install guide.
• See the examples gallery providing different inference workflow samples.
• Swap out data sources or models depending on your use case!

Tutorial

Agent-assisted setup

Automate setup with your preferred coding agent using NVIDIA Earth2Studio skills. Install the Earth2Studio skill set, then ask your favorite agent (Claude, Codex, OpenCode, etc) to recommend a model, configure an environment, or run a first deterministic forecast. Find more Earth2Studio skills in the NVIDIA Skills catalog.

npx skills add NVIDIA/skills --skill earth2studio-install
npx skills add NVIDIA/skills --skill earth2studio-discover
npx skills add NVIDIA/skills --skill earth2studio-data-fetch
npx skills add NVIDIA/skills --skill earth2studio-deterministic-forecast

Example agent prompts:

Use the Earth2Studio discover skill to recommend a starter forecast workflow.
Use the Earth2Studio install skill to set up my environment for FourCastNet3 inference.
Create a script to fetch ERA5 surface winds data for March 2024.
Create a deterministic forecast workflow with GFS, FourCastNet3, and a Zarr output store.

NVIDIA FourCastNet3

from earth2studio.models.px import FCN3
from earth2studio.data import GFS
from earth2studio.io import ZarrBackend
from earth2studio.run import deterministic as run

model = FCN3.load_model(FCN3.load_default_package())
data = GFS()
io = ZarrBackend("outputs/fcn3_forecast.zarr")
run(["2025-01-01T00:00:00"], 10, model, data, io)

ECMWF AIFS

from earth2studio.models.px import AIFS
from earth2studio.data import IFS
from earth2studio.io import ZarrBackend
from earth2studio.run import deterministic as run

model = AIFS.load_model(AIFS.load_default_package())
data = IFS()
io = ZarrBackend("outputs/aifs_forecast.zarr")
run(["2025-01-01T00:00:00"], 10, model, data, io)

Google Graphcast

from earth2studio.models.px import GraphCastOperational
from earth2studio.data import GFS
from earth2studio.io import ZarrBackend
from earth2studio.run import deterministic as run

package = GraphCastOperational.load_default_package()
model = GraphCastOperational.load_model(package)
data = GFS()
io = ZarrBackend("outputs/graphcast_operational_forecast.zarr")
run(["2025-01-01T00:00:00"], 4, model, data, io)

[!IMPORTANT] Earth2Studio is an interface to third‑party models, checkpoints, and datasets. Licenses for these assets are owned by their providers. Ensure you have the rights to download, use, and (if applicable) redistribute each model and dataset. Links to the original license and source are often provided in the API docs for each model/data source.

Latest News

[!NOTE] As of version 0.14.0, Earth2Studio TOML default installs now target CUDA 13.

• Himawari AHI, Himawari-8/9 AHI ISatSS L2 Full Disk satellite data source is now available for geostationary satellite observations over the Western Pacific.
• GHCN-Daily, Global Historical Climatology Network daily station observation data frame source providing access to global surface observations.
• Orbit-2, precipitation downscaling diagnostic model added for high-resolution precipitation estimates.
• Conventional Observation Sources, new data sources for in-situ observations including NNJA conventional obs and real-time GDAS conventional obs.

For a complete list of latest features and improvements see the changelog.

Overview

Earth2Studio is an AI inference pipeline toolkit focused on weather and climate applications that is designed to ride on top of different AI frameworks, model architectures, data sources and SciML tooling while providing a unified API.

The composability of the different core components in Earth2Studio easily allows the development and deployment of increasingly complex pipelines that may chain multiple data sources, AI models and other modules together.

The unified ecosystem of Earth2Studio provides users the opportunity to rapidly swap out components for alternatives. In addition to the largest model zoo of weather/climate AI models, Earth2Studio is packed with useful functionality such as optimized data access to cloud data stores, statistical operations and more to accelerate your pipelines.

Earth-2 Open Models

Access state of the art Nvidia open models for climate and weather: Earth-2 Open Models. For training recipes for these models, see the PhysicsNeMo repository.

Features

Earth2Studio package focuses on supplying you the tools to build your own workflows, pipelines, APIs, or packages using modular components including:

Prognostic Models

Prognostic models in Earth2Studio perform time integration, taking atmospheric fields at a specific time and auto-regressively predicting the same fields into the future (typically 6 hours per step), enabling both single time-step predictions and extended time-series forecasting.

Earth2Studio maintains the largest collection of pre-trained state-of-the-art AI weather/climate models ranging from global forecast models to regional specialized models, covering various resolutions, architectures, and forecasting capabilities to suit different computational and accuracy requirements.

Available models include but are not limited to:

Model	Resolution	Architecture	Time Step	Coverage
GraphCast Small	1.0°	Graph Neural Network	6h	Global
GraphCast Operational	0.25°	Graph Neural Network	6h	Global
Pangu 3hr	0.25°	Transformer	3h	Global
Pangu 6hr	0.25°	Transformer	6h	Global
Pangu 24hr	0.25°	Transformer	24h	Global
Aurora	0.25°	Transformer	6h	Global
FuXi	0.25°	Transformer	6h	Global
AIFS	0.25°	Transformer	6h	Global
AIFS Ensemble	0.25°	Transformer Ensemble	6h	Global
StormCast	3km	Diffusion + Regression	1h	Regional (US)
SFNO	0.25°	Neural Operator	6h	Global
DLESyM	0.25°	Convolutional	6h	Global

For a complete list, see the prognostic model API docs.

Diagnostic Models

Diagnostic models in Earth2Studio perform time-independent transformations, typically taking geospatial fields at a specific time and predicting new derived quantities without performing time integration enabling users to build pipelines to predict specific quantities of interest that may not be provided by forecasting models.

Earth2Studio contains a growing collection of specialized diagnostic models for various phenomena including precipitation prediction, tropical cyclone tracking, solar radiation estimation, wind gust forecasting, and more.

Available diagnostics include but are not limited to:

Model	Resolution	Architecture	Coverage	Output
PrecipitationAFNO	0.25°	Neural Operator	Global	Total precipitation
SolarRadiationAFNO1H	0.25°	Neural Operator	Global	Surface solar radiation
WindgustAFNO	0.25°	AFNO	Global	Maximum wind gust
TCTrackerVitart	0.25°	Algorithmic	Global	TC tracks & properties
CBottleInfill	100km	Diffusion	Global	Global climate sample
CBottleSR	5km	Diffusion	Regional / Global	High-res climate
CorrDiff	Variable	Diffusion	Regional	Fine-scale weather
CorrDiffTaiwan	2km	Diffusion	Regional (Taiwan)	Taiwan fine-scale weather

For a complete list, see the diagnostic model API docs.

Datasources

Data sources in Earth2Studio provide a standardized API for accessing weather and climate datasets from various providers (numerical models, data assimilation results, and AI-generated data), enabling seamless integration of initial conditions for model inference and validation data for scoring across different data formats and storage systems.

Earth2Studio includes data sources ranging from operational weather models (GFS, HRRR, IFS) and reanalysis datasets (ERA5 via ARCO, CDS) to AI-generated climate data (cBottle) and local file systems. Fetching data is just plain easy, Earth2Studio handles the complicated parts giving you an easy to use Xarray data array of requested data under a shared package wide vocabulary and coordinate system.

Available data sources include but are not limited to:

Data Source	Type	Resolution	Coverage	Data Format
GFS	Operational	0.25°	Global	GRIB2
GFS_FX	Forecast	0.25°	Global	GRIB2
HRRR	Operational	3km	Regional (US)	GRIB2
HRRR_FX	Forecast	3km	Regional (US)	GRIB2
ARCO ERA5	Reanalysis	0.25°	Global	Zarr
CDS	Reanalysis	0.25°	Global	NetCDF
IFS	Operational	0.25°	Global	GRIB2
NCAR_ERA5	Reanalysis	0.25°	Global	NetCDF
WeatherBench2	Reanalysis	0.25°	Global	Zarr
GEFS_FX	Ensemble Forecast	0.25°	Global	GRIB2
ISD	Observational	Point	Regional (US)	CSV
MRMS	Reanalysis	1km	Regional (US)	GRIB2

For a complete list, see the data source API docs.

IO Backends

IO backends in Earth2Studio provides a standardized interface for writing and storing pipeline outputs across different file formats and storage systems enabling users to store inference outputs for later processing.

Earth2Studio includes IO backends ranging from traditional scientific formats (NetCDF) and modern cloud-optimized formats (Zarr) to in-memory storage backends.

Available IO backends include:

IO Backend	Format	Features	Location
ZarrBackend	Zarr	Compression, Chunking	In-Memory/Local
AsyncZarrBackend	Zarr	Async writes, Parallel I/O	In-Memory/Local/Remote
NetCDF4Backend	NetCDF4	CF-compliant, Metadata	In-Memory/Local
XarrayBackend	Xarray Dataset	Rich metadata, Analysis-ready	In-Memory
KVBackend	Key-Value	Fast Temporary Access	In-Memory

For a complete list, see the IO API docs.

Perturbation Methods

Perturbation methods in Earth2Studio provide a standardized interface for adding noise to data arrays, typically enabling the creation of ensembling forecast pipelines that capture uncertainty in weather and climate predictions.

Available perturbations include but are not limited to:

Perturbation Method	Type	Spatial Correlation	Temporal Correlation
Gaussian	Noise	None	None
Correlated SphericalGaussian	Noise	Spherical	AR(1) process
Spherical Gaussian	Noise	Spherical (Matern)	None
Brown	Noise	2D Fourier	None
Bred Vector	Dynamical	Model-dependent	Model-dependent
Hemispheric Centred Bred Vector	Dynamical	Hemispheric	Model-dependent

For a complete list, see the perturbations API docs.

Statistics / Metrics

Statistics and metrics in Earth2Studio provide operations typically useful for in-pipeline evaluation of forecast performance across different dimensions (spatial, temporal, ensemble) through various statistical measures including error metrics, correlation coefficients, and ensemble verification statistics.

Available operations include but are not limited to:

Statistic	Type	Application
RMSE	Error Metric	Forecast accuracy
ACC	Correlation	Pattern correlation
CRPS	Ensemble Metric	Probabilistic skill
Rank Histogram	Ensemble Metric	Ensemble reliability
Standard Deviation	Moment	Spread measure
Spread-Skill Ratio	Ensemble Metric	Ensemble calibration

For a complete list, see the statistics API docs.

For a more complete list of features, be sure to view the documentation. Don't see what you need? Great news, extension and customization are at the heart of our design.

Contributors

Check out the contributing document for details about the technical requirements and the user guide for higher level philosophy, structure, and design.

License

Earth2Studio is provided under the Apache License 2.0, refer to the LICENSE file for full license text.