PhysicsNeMo Checkpoints: StormScope-Meteosat

Description:

StormScope-Meteosat is a mesoscale weather AI model that autoregressively predicts Meteosat Third Generation (MTG) Flexible Combined Imager (FCI) satellite variables. The model can make predictions at the native spatiotemporal resolution of the FCI infrared channels (2 km at the sub-satellite point, 10 minutes).

For inference see NVIDIA Earth2Studio.

This model is ready for commercial or non-commercial use.

License/Terms of Use:

Governing Terms: Use of this model is governed by the Linux Foundation OpenMDW License Agreement, version 1.1. See LICENSE file for governing terms.

Deployment Geography:

Global

Use Case:

High-resolution short-term ensemble forecasting of Meteosat FCI observations

Release Date:

Hugging Face: 07/27/2026 via https://huggingface.co/nvidia/stormscope-meteosat

References:

• Stormscope US paper: https://arxiv.org/abs/2601.17268
• MTG FCI level 1c data guide: https://user.eumetsat.int/resources/user-guides/mtg-fci-level-1c-data-guide

Model Architecture:

Architecture Type: StormScope-Meteosat uses a diffusion transformer architecture with 2D neighborhood attention.
Network Architecture: Diffusion Transformer (DiT)

This model provides the following checkpoints:

• European grid (1120x2432), DiT -- 306M Parameters
• EMEA grid (4672x4160), DiT -- 176M Parameters

Input(s):

Input Type(s):

• Tensor: (6 time steps of 16 state variables from Meteosat FCI and 3 solar geometry variables), as well as 8 conditioning variables.
• DateTime (NumPy Array)
  

Input Format(s): PyTorch Tensor / NumPy array
Input Parameters:

• Five Dimensional (5D) (batch, lead time, variable, elevation, azimuth)
  
• Input DateTime (1D)
  

Other Properties Related to Input:

• Input grid (height/width) dimensions correspond to the native FCI grid with 2 km resolution at the sub-satellite point.
• Input lead time is of size 6 (model uses 6 time steps as input)
• Input state FCI variables: fci_vis_04, fci_vis_05, fci_vis_06, fci_vis_08, fci_vis_09, fci_nir_13, fci_nir_16, fci_nir_22, fci_ir_38, fci_wv_63, fci_wv_73, fci_ir_87, fci_ir_97, fci_ir_105, fci_ir_123, fci_ir_133
• Conditioning variables:
  • Digital elevation and water body models (ASTER):
    • mean surface elevation of pixel
    • standard deviation of surface elevation within pixel
    • fraction of land surface (rather than water) in pixel
  • FCI time offset: the approximate difference of FCI scan time from the reference time
  • Solar geometry (computed):
    • cosine of solar zenith angle
    • sine and cosine of solar azimuth angle
  • Satellite viewing geometry (computed):
    • viewing azimuth and zenith angles w.r.t. sub-satellite point
  • Earth surface geometry (computed):
    • along-surface displacement from sub-satellite point in north-south and east-west directions
    • cosine of satellite elevation angle from surface

For variable naming information, review the FCI Lexicon in Earth2Studio.

Output:

Output Type(s): Tensor: 16 state variables from Meteosat FCI.
Output Format: PyTorch Tensors
Output Parameters: Five Dimensional (5D) (batch, lead time, variable, elevation, azimuth)
Other Properties Related to Output:

• Output grid (height/width) dimensions correspond to the native FCI grid with 2 km resolution at the sub-satellite point.
• Output lead time is of size 1 (model predicts one timestep)
• Output state FCI variables: same as input state FCI variables

Our AI models are designed and/or optimized to run on NVIDIA GPU-accelerated systems. By leveraging NVIDIA's hardware (e.g. GPU cores) and software frameworks (e.g., CUDA libraries), the model achieves faster training and inference times compared to CPU-only solutions.

Software Integration:

Runtime Engine(s): Not Applicable
Supported Hardware Microarchitecture Compatibility:

• NVIDIA Ampere
  
• NVIDIA Blackwell
  
• NVIDIA Hopper
  

Supported Operating System(s):

• Linux
  

The integration of foundation and fine-tuned models into AI systems requires additional testing using use-case-specific data to ensure safe and effective deployment. Following the V-model methodology, iterative testing and validation at both unit and system levels are essential to mitigate risks, meet technical and functional requirements, and ensure compliance with safety and ethical standards before deployment.

Model Version(s):

Model Version: v1

Training, Testing, and Evaluation Datasets:

Training Dataset:

Link: Meteosat FCI

Data Collection Method by dataset:

• Automatic/Sensors
  

Labeling Method by dataset:

• Automatic/Sensors
  

Data Modality:

• Gridded geophysical time series

Data Size:

• 25 TB processed and compressed subset used for model training

Properties: Meteosat Third Generation FCI imaging data for the period November 2024 - May 2026. FCI is a EUMETSAT geostationary satellite instrument aboard MTG providing multispectral radiances at 0.5 km - 2 km resolution over a full Earth disk.

Contains modified EUMETSAT Meteosat FCI Level 1c Normal Resolution Image Data (2024-2026)

Link: ASTER GDEM/ASTWBD

Data Collection Method by dataset:

• Automatic/Sensors
  

Labeling Method by dataset:

• Automatic/Sensors
  

Data Modality:

• Gridded geophysical static data

Data Size:

• 496 MB downsampled subset used for model training

Properties: Global digital elevation model (DEM) and water body map. Native resolution ~30 m/pixel, downsampled to target domain grid.

Testing/Evaluation Dataset:

Link: Meteosat FCI

Data Collection Method by dataset:

• Automatic/Sensors
  

Labeling Method by dataset:

• Automatic/Sensors
  

Properties: Meteosat Third Generation FCI imaging data for October 2024. FCI is a EUMETSAT geostationary satellite instrument aboard MTG providing multispectral radiances at 0.5 km - 2 km resolution over a full Earth disk.

Contains modified EUMETSAT Meteosat FCI Level 1c Normal Resolution Image Data (2024)

Link: ASTER GDEM/ASTWBD

Data Collection Method by dataset:

• Automatic/Sensors
  

Labeling Method by dataset:

• Automatic/Sensors
  

Properties: Global digital elevation model (DEM) and water body map. Native resolution ~30 m/pixel, downsampled to target domain grid.

Inference:

Acceleration Engine: PyTorch
Test Hardware:

• NVIDIA Hopper H100
  
• NVIDIA Blackwell B200
  
• NVIDIA Blackwell B300
  

Ethical Considerations:

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. Developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.

For more detailed information on ethical considerations for this model, please see the Model Card++ Explainability, Bias, Safety & Security, and Privacy Subcards.

Please report model quality, risk, security vulnerabilities or NVIDIA AI Concerns here.