PhysicsNeMo Checkpoints: StormCast-CONUS

Description:

StormCast-CONUS is a high-resolution generative regional weather forecasting model that autoregressively predicts 99 state variables at km scale over the Continental United States (CONUS), using a 1-hour time step with dense vertical resolution in the atmospheric boundary layer. StormCast-CONUS is a further development of StormCast V1, replacing the UNet backbone with a Diffusion Transformer (DiT) architecture that uses neighborhood attention (NATTEN) for efficient high-resolution attention computation.

For training recipes see NVIDIA PhysicsNeMo, 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.

Deployment Geography:

Global

Use Case:

High-resolution regional weather forecasting and ensemble generation over CONUS at km scale

Reference(s):

• StormCast v1: Kilometer-Scale Convection Allowing Model Emulation using Generative Diffusion Modeling
  

Model Architecture:

Architecture Type: StormCast-CONUS uses a Diffusion Transformer (DiT) architecture with neighborhood attention (NATTEN), 265M parameters.
Network Architecture: Diffusion Transformer (DiT) with Neighborhood Attention (NATTEN)

Input:

Input Type(s):

• Tensor (125 Surface and Model level variables - 99 state variables and 26 conditioning variables.)
• DateTime (NumPy Array)
  

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

• Four Dimensional (4D) (batch, variable, latitude, longitude)
  
• Input DateTime (1D)
  

Other Properties Related to Input:

• Input latitude/longitude grid: HRRR Lambert Conformal Conic projection, 3 km resolution, 1024×1792 grid
• Input state weather variables: u10m, v10m, t2m, mslp, u_hl1, u_hl2, u_hl3, u_hl4, u_hl5, u_hl6, u_hl7, u_hl8, u_hl9, u_hl10, u_hl11, u_hl13, u_hl15, u_hl20, u_hl25, u_hl30, v_hl1, v_hl2, v_hl3, v_hl4, v_hl5, v_hl6, v_hl7, v_hl8, v_hl9, v_hl10, v_hl11, v_hl13, v_hl15, v_hl20, v_hl25, v_hl30, t_hl1, t_hl2, t_hl3, t_hl4, t_hl5, t_hl6, t_hl7, t_hl8, t_hl9, t_hl10, t_hl11, t_hl13, t_hl15, t_hl20, t_hl25, t_hl30, q_hl1, q_hl2, q_hl3, q_hl4, q_hl5, q_hl6, q_hl7, q_hl8, q_hl9, q_hl10, q_hl11, q_hl13, q_hl15, q_hl20, q_hl25, q_hl30, z_hl1, z_hl2, z_hl3, z_hl4, z_hl5, z_hl6, z_hl7, z_hl8, z_hl9, z_hl10, z_hl11, z_hl13, z_hl15, z_hl20, z_hl25, z_hl30, p_hl1, p_hl2, p_hl3, p_hl4, p_hl5, p_hl6, p_hl7, p_hl8, p_hl9, p_hl10, p_hl11, p_hl13, p_hl15, p_hl20, refc
  
• Conditioning weather variables: u10m, v10m, t2m, tcwv, mslp, sp, u1000, u850, u500, u250, v1000, v850, v500, v250, z1000, z850, z500, z250, t1000, t850, t500, t250, q1000, q850, q500, q250
• Cosine of solar zenith angle (computed)
• Constant inputs: mean and standard deviation of surface elevation in pixel; land/water mask

For variable naming information, review the HRRR Lexicon at Earth2Studio, but u, v, t, z and p refer to winds, temperature, geopotential, and pressure (respectively). Variables marked with _hl refer to natural/hybrid model levels.

Output:

Output Type(s): Tensor (99 Surface and Model level variables)
Output Format: PyTorch Tensors
Output Parameters: Four Dimensional (4D) (batch, variable, latitude, longitude)
Other Properties Related to Output:

• Output latitude/longitude grid: HRRR Lambert Conformal Conic projection, 3 km resolution, 1024×1792 grid
• Output state weather variables: u10m, v10m, t2m, mslp, u_hl1, u_hl2, u_hl3, u_hl4, u_hl5, u_hl6, u_hl7, u_hl8, u_hl9, u_hl10, u_hl11, u_hl13, u_hl15, u_hl20, u_hl25, u_hl30, v_hl1, v_hl2, v_hl3, v_hl4, v_hl5, v_hl6, v_hl7, v_hl8, v_hl9, v_hl10, v_hl11, v_hl13, v_hl15, v_hl20, v_hl25, v_hl30, t_hl1, t_hl2, t_hl3, t_hl4, t_hl5, t_hl6, t_hl7, t_hl8, t_hl9, t_hl10, t_hl11, t_hl13, t_hl15, t_hl20, t_hl25, t_hl30, q_hl1, q_hl2, q_hl3, q_hl4, q_hl5, q_hl6, q_hl7, q_hl8, q_hl9, q_hl10, q_hl11, q_hl13, q_hl15, q_hl20, q_hl25, q_hl30, z_hl1, z_hl2, z_hl3, z_hl4, z_hl5, z_hl6, z_hl7, z_hl8, z_hl9, z_hl10, z_hl11, z_hl13, z_hl15, z_hl20, z_hl25, z_hl30, p_hl1, p_hl2, p_hl3, p_hl4, p_hl5, p_hl6, p_hl7, p_hl8, p_hl9, p_hl10, p_hl11, p_hl13, p_hl15, p_hl20, refc
  

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 Hopper
  
• NVIDIA Blackwell
  

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: 1.0.0

Training, Testing, and Evaluation Datasets:

Training Dataset:

Link: ERA5

Data Collection Method by dataset:

• Automatic/Sensors
  

Labeling Method by dataset:

• Automatic/Sensors
  

Data Modality:

• Gridded geophysical time series

Data Size:

• 150 GB subset used for model training

Properties: ERA5 data for the period 2018-07-12 – 2025-09-30. ERA5 provides hourly estimates of various atmospheric, land, and oceanic climate variables. The data covers the Earth on a 30km grid and resolves the atmosphere at 137 levels.

Link: HRRR

Data Collection Method by dataset:

• Automatic/Sensors
  

Labeling Method by dataset:

• Automatic/Sensors
  

Data Modality:

• Gridded geophysical time series

Data Size:

• 15 TB subset used for model training

Properties: HRRR data for the date range of 2018-07-12 to 2025-09-30. The HRRR is a NOAA real-time 3-km resolution, hourly updated, cloud-resolving, convection-allowing atmospheric model, initialized by 3 km grids with 3 km radar assimilation.

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:

• 120 MB upsampled subset used for model training

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

Testing/Evaluation Dataset:

Link: ERA5

Data Collection Method by dataset:

• Automatic/Sensors
  

Labeling Method by dataset:

• Automatic/Sensors
  

Properties: ERA5 data for the date range of 2025-10-01 - 2025-12-31. ERA5 provides hourly estimates of various atmospheric, land, and oceanic climate variables. The data covers the Earth on a 30km grid and resolves the atmosphere at 137 levels.

Link: HRRR

Data Collection Method by dataset:

• Automatic/Sensors
  

Labeling Method by dataset:

• Automatic/Sensors
  

Properties: HRRR data for the date range of 2025-10-01 - 2025-12-31. The HRRR is a NOAA real-time 3-km resolution, hourly updated, cloud-resolving, convection-allowing atmospheric model, initialized by 3 km grids with 3 km radar assimilation.

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, upsampled to HRRR grid.

Inference:

Engine: PyTorch / Earth2Studio
Test Hardware:

• NVIDIA Ampere (A100)
  
• NVIDIA Hopper (H100)
  
• NVIDIA Ada Lovelace (L40S)
  

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.