STRATA

STRATA

Emulates a kilometer-scale global storm-resolving atmosphere simulation with a 10-minute time step, for weather and climate research. An NVIDIA Earth-2 checkpoint with 331.7M parameters.

Made by nvidia on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: STRATA is an NVIDIA Earth-2 checkpoint, an AI atmosphere model that imitates the SCREAM global storm-resolving simulation. It predicts the three-dimensional state of the atmosphere over the whole globe at about 4.9 km grid spacing, one 10-minute step at a time.

The state includes potential temperature, winds, geopotential height, vertical pressure velocity, specific humidity and 2-m air temperature. Surface pressure and precipitation fluxes are worked out at every step.

Who it suits: Weather and climate researchers who want a fast emulator of a storm-resolving simulation, and groups that study how AI models of different resolutions can work together.

Industry | Who uses it | How they use it | Example
Climate science | Atmospheric researchers | Emulate a storm-resolving simulation | Advance the SCREAM state step by step
Weather research | Forecast researchers | Run short-range forecasts of the SCREAM state | Look hours to days ahead
Research computing | High-performance computing teams | Split the globe across several GPUs | Run the full globe with domain decomposition
Machine learning | Teams studying coupled emulators | Test nudging or downscaling between resolutions | Combine emulators of different resolutions

Skip it if you need a ready-made forecast service, since this is a research checkpoint that needs code and data preparation.

Where to use it:

• A single GPU, which can advance any part of the globe
• Several GPUs, which the card says advance the full globe through domain decomposition
• The public STRATA project on GitHub, which holds the inference wrapper and training recipes
• NVIDIA PhysicsNeMo, which holds the network architecture
• Windows, macOS or Linux with Python, to download the files

Use case | Who | What they get back
Short-range forecasting | Weather researchers | The SCREAM state hours to days ahead
Storm-resolving emulation | Climate researchers | A fast stand-in for a heavy simulation
Coupling experiments | Machine learning researchers | A resolution to nudge or downscale against

How to set it up:

• Python and pip to run the download commands
• One GPU at least, with several for the full globe
• The STRATA inference wrapper from GitHub and the PhysicsNeMo model code
• Input data on the SCREAM ne1024pg2 grid

1. Install the Hugging Face tools.
2. Download the files, which include the checkpoint and three static field files.
3. Set up the STRATA inference wrapper and PhysicsNeMo as the card describes.
4. Prepare your input tensor and run the model.

pip install -U huggingface_hub
hf download nvidia/strata

How to use it: Prepare an input tensor of 149 channels, made of 145 state channels and 4 forcing channels. Each call returns the state 10 minutes later, so you repeat it to look further ahead.

Item | Detail
Grid | SCREAM ne1024pg2 cubed sphere, 6 faces of 2048 by 2048 columns
Time step | 10 minutes
Input | 149 channels, normalized with statistics stored in the checkpoint
Output | 148 channels, the 145 state channels plus 3 diagnostic channels
Forcing channels | The cosine of the solar zenith angle, surface geopotential, orography spread and land fraction

Example: You feed in the current global state with its forcing fields and run one step. The model returns the state 10 minutes later, and feeding that back in several times advances the forecast.

For newcomers and experts: If you are new to weather models, think of this as a fast copy of a huge climate simulation. Download the files, read the card, and plan on running the example with the GitHub wrapper.

If you know the field, it is a two-stage transformer with a semantic backbone on 4 by 4 pixel patches and a pixel-resolution refinement stage. It predicts each step in a single deterministic pass and trains with a regression loss, so it does not sample like a diffusion model.

Global state and forcing channels
Tiled inference over cubed-sphere tiles
Two-stage transformer predicts one step
State 10 minutes later

Good to know:

• Inference is tile based, with overlapping tiles of 128 to 256 pixels plus halos
• The checkpoint went through pretraining and a 4-step autoregressive fine-tune
• The three static fields come from datasets under CC BY 4.0, with attribution notes in the repository

Title: Channels in the model tensors
Input channels: 149 channels
State channels: 145 channels
Forcing channels: 4 channels
Output channels: 148 channels
Diagnostic channels: 3 channels

• License: OpenMDW 1.1, and the card allows both commercial and non-commercial use
• Cost to run: none named, though full-globe runs need several GPUs
• Data it reaches: the atmospheric state tensors and static field files you provide, plus the files you download
• Not for: other grids or time steps, since the checkpoint is made for the ne1024pg2 grid and a 10-minute step

On Hugging Face, 13 November 2026

• Downloads in the last 30 days: 249
• Likes: 27
• License: openmdw-1.1
• Tags: pixeldit

Links

• Model: https://huggingface.co/nvidia/strata
• Files: https://huggingface.co/nvidia/strata/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download nvidia/strata

Then follow the usage in its model card. Nothing in this kit runs by itself.