| Field | Response |
|---|---|
| Intended Task/Domain: | Automotive Aerodynamics |
| Model Type: | Transformer with Geometry-Aware Latent Embeddings (GeoTransolver) |
| Intended Users: | Automotive engineers and researchers accelerating Computational Fluid Dynamics (CFD) predictions using AI. |
| Output: | Tensor (surface pressure and wall shear stress; volumetric velocity and pressure fields). |
| Describe how the model works: | GeoTransolver uses GALE (Geometry-Aware Latent Embedding) attention blocks that combine physics-aware self-attention over learned physical state slices (from Transolver) with cross-attention to a shared multi-scale geometry context. The geometry context is constructed once via multi-scale ball queries at 6 radii and encodes local surface geometry attributes and global boundary conditions. A learnable gate blends self-attention and cross-attention outputs in each layer, enabling the model to adaptively weight physics-state interactions vs. geometry conditioning. |
| Name the adversely impacted groups this has been tested to deliver comparable outcomes regardless of: | Not Applicable |
| Technical Limitations & Mitigation: | Increased errors are observed near complex flow regions such as wheels, mirrors, separation zones, and wake regions with sharp gradients. Lift prediction may show oscillations in configurations with sparser training data coverage in certain operating regimes. |
| Verified to have met prescribed NVIDIA quality standards: | Yes |
| Performance Metrics: | Surface and volume prediction relative L1 error; on DrivAerML (48 test designs): surface pressure 2.86%, wall shear 4.9%, volume pressure 3.09%, volume velocity 4.02%, drag R² 0.996, lift R² 0.991 |
| Potential Known Risks: | This model may inaccurately predict aerodynamic fields for vehicle designs outside the training distribution. |
| Licensing: | Use of this model is governed by the NVIDIA Open Model Agreement. |