nvidia/geotransolver_drivaerml

GeoTransolver DrivAerML

GeoTransolver DrivAerML is a Transformer-based surrogate model for large-scale automotive external aerodynamics simulations. It extends Physics-Attention with Geometry-Aware Latent Embeddings (GALE), coupling learnable physical state slice self-attention with cross-attention to a shared multi-scale geometry and boundary condition context. The model predicts surface pressure and wall shear stress fields, as well as volumetric velocity and pressure fields on 3D vehicle geometries for computational fluid dynamics (CFD) applications.

This model is available for commercial use.

License/Terms of Use:

Use of this model is governed by the NVIDIA Open Model Agreement.

Deployment Geography:

Global

Use Case:

Computational Fluid Dynamics (CFD) engineers accelerating automotive external aerodynamics with AI.

Release Date:

05/01/2026

Hugging Face: https://huggingface.co/nvidia/geotransolver_drivaerml

Reference(s):

Code

GeoTransolver: Learning Physics on Irregular Domains using Multi-scale Geometry Aware Physics Attention Transformer

DrivAerML: High-Fidelity Computational Fluid Dynamics Dataset for Road-Car External Aerodynamics

Model Architecture:

Architecture Type: Transformer with Geometry-Aware Latent Embeddings (GALE) combining physics-aware self-attention and multi-scale geometry cross-attention.

Network Architecture: GeoTransolver is built on GALE attention blocks, each comprising: (1) Physics-Aware Self-Attention that learns soft assignments of input points to M latent physical state slices (inheriting from Transolver), with slice-wise self-attention via Q/K/V projections; (2) Cross-Attention to a shared geometry context vector, which encodes multi-scale local geometry features extracted via ball queries at 6 radii (0.01–5.0) using k=32 nearest neighbors, followed by MLP processing, mean/max/attention pooling, and concatenation with global boundary condition parameters; and (3) an adaptive gate (learnable sigmoid parameter) that blends self-attention and cross-attention outputs. The geometry context is computed once and shared across all layers.

Number of model parameters: 29M (20 GALE layers, six-scale ball-query radii, kernel size 32)

Input:

Input Type(s):

• Tensor (3D point cloud coordinates on vehicle surface and volume, plus global boundary condition parameters)

Input Format(s): PyTorch Tensor
Input Parameters:

• Surface geometry: point coordinates (M_g, 3) and attributes including normals and curvatures (M_g, d_g)
• Input slices: 3D positions (N_m, 3) with d_x-dimensional features for surface and volume points
• Global parameters: boundary conditions and operating regime (d_p,)

Other Properties Related to Input:

• Multi-scale geometry context is computed once at 6 spatial radii (0.01, 0.05, 0.25, 1.0, 2.5, 5.0) and shared across all layers
• Coordinates normalized to the vehicle bounding box

Output:

Output Type(s): Tensor (Surface and volume aerodynamic fields)
Output Format: PyTorch Tensor
Output Parameters:

• Surface: pressure (M_s, 1), wall shear stress (M_s, 3)
• Volume: velocity (M_v, 3), pressure (M_v, 1)

Other Properties Related to Output:

• Outputs are normalized using statistics computed from the training dataset
• Drag and lift coefficients derived via surface integration of pressure and wall shear stress predictions

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): PyTorch
Supported Hardware Microarchitecture Compatibility:

• NVIDIA Ampere
  
• NVIDIA Blackwell
  
• NVIDIA Hopper
  
• NVIDIA Turing
  

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:

The DrivAerML dataset is used for training and evaluation, which is a publicly available, high-fidelity dataset comprising aerodynamic data for 500 parametrically morphed variants of the DrivAer notchback vehicle. The dataset was generated using hybrid RANS/LES (HRLES), a scale-resolving CFD method, which provides time-averaged quantities for each variant. The available data includes surface pressure, wall shear stress, and flow-field quantities, provided in formats compatible with mesh-based analysis (.vtp for surface data and .vtu for flow-field data). 48 samples (~10%) are used as the test set, with approximately 20% of the test set consisting of out-of-distribution samples based on drag coefficients. These samples represent extreme cases with the lowest and highest drag coefficients in the entire dataset, which remain unseen by the model during training. Models are trained for up to 500 epochs on a single NVIDIA GB200 node using the Muon optimizer.

Training Dataset:

Data Modality:

• Other: 3D Point Cloud (surface and volume)

Training Data Size:

• 436 files in VTP format (surface meshes) and VTU format (volume flow fields) with corresponding physical quantities

Link: DrivAerML Dataset

Data Collection Method by dataset:

• Synthetic CFD Simulation
  

Labeling Method by dataset:

• Automated
  

Properties: The data is a simulation/synthetic dataset generated using hybrid RANS/LES scale-resolving CFD simulations, providing time-averaged surface and volumetric flow fields for different car geometries. Each case contains approximately 150 million volume elements and 10 million surface elements.

Testing Dataset:

Link: DrivAerML Dataset

Data Collection Method by dataset:

• Synthetic CFD Simulation
  

Labeling Method by dataset:

• Synthetic CFD Simulation
  

Properties: Test split from DrivAerML dataset with vehicle geometries held out from training. 48 samples (~10%) are used as the test set, with approximately 20% consisting of out-of-distribution samples based on drag coefficients.

Evaluation Dataset:

Link: DrivAerML Dataset

Data Collection Method by dataset:

• Synthetic CFD Simulation
  

Labeling Method by dataset:

• Automated
  

Properties: Validation split from DrivAerML dataset with vehicle geometries held out from training. The full DrivAerML dataset is split as 90% for training and 10% for validation.

Inference:

Acceleration Engine: PyTorch
Test Hardware:

• H100
  
• GB200
  

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. When downloaded or used in accordance with our terms of service, 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++ subcards: Bias, Explainability, Privacy, and Safety & Security.

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