PaperlayerView exact revision on arXiv

Source-linked AI summary

TripNet: Learning Large-scale High-fidelity 3D Car Aerodynamics with Triplane Networks

Qian Chen, Mohamed Elrefaie, Angela Dai, Faez Ahmed

arXiv:2503.17400v2physics.flu-dyncs.LG

TL;DR

Large-scale CFD surrogate models are limited by explicit, mesh-dependent representations that become memory-intensive and resolution-limited for simulations with millions of nodes and cells. TripNet instead encodes car geometry in compact triplanes and queries aerodynamic quantities at arbitrary spatial locations. It achieves state-of-the-art results across drag, surface fields, and 3D flow prediction on DrivAerNet and DrivAerNet++ while improving computational efficiency.

  • Problem

    Existing CFD surrogate models rely on explicit geometric representations that are memory-intensive and resolution-limited for large-scale simulations.

  • Method

    TripNet implicitly encodes each 3D car geometry into compact triplane features and uses them for query-based aerodynamic prediction across multiple tasks.

  • Results

    TripNet achieves state-of-the-art performance on drag, surface pressure, wall shear stress, and 3D flow prediction across DrivAerNet and DrivAerNet++.

  • Takeaways & Limitations

    TripNet offers a scalable and efficient alternative to traditional CFD solvers and existing surrogate models for high-fidelity aerodynamic analysis.

  • Takeaways & Limitations

    Generalization to topologically distinct shapes and flow regimes beyond the training Reynolds number remains unverified, and physical constraints are not explicitly enforced.

Abstract

from arXiv · show

Surrogate modeling has emerged as a powerful tool to accelerate Computational Fluid Dynamics (CFD) simulations. Existing 3D geometric learning models based on point clouds, voxels, meshes, or graphs depend on explicit geometric representations that are memory-intensive and resolution-limited. For large-scale simulations with millions of nodes and cells, existing models require aggressive downsampling due to their dependence on mesh resolution, resulting in degraded accuracy. We present TripNet, a triplane-based neural framework that implicitly encodes 3D geometry into a compact, continuous feature map with fixed dimension. Unlike mesh-dependent approaches, TripNet scales to high-resolution simulations without increasing memory cost, and enables CFD predictions at arbitrary spatial locations in a query-based fashion, independent of mesh connectivity or predefined nodes. TripNet achieves state-of-the-art performance on the DrivAerNet and DrivAerNet++ datasets, accurately predicting drag coefficients, surface pressure, and full 3D flow fields. With a unified triplane backbone supporting multiple simulation tasks, TripNet offers a scalable, accurate, and efficient alternative to traditional CFD solvers and existing surrogate models.

1 Introduction

TripNet addresses the cost and resolution limits of large-scale CFD surrogate modeling with a compact triplane representation that supports arbitrary spatial queries. It achieves fast, accurate predictions across drag, surface fields, and 3D flow tasks while remaining scalable across aerodynamic simulations.

  • The method targets the computational burden of high-fidelity simulations, which can require days or weeks and substantial resources.Existing mesh-based and graph-based approaches depend on discrete simulation representations and message passing over mesh edges.
  • TripNets encode aerodynamic simulations in triplanes, enabling solutions to be queried at arbitrary 3D locations rather than only at mesh nodes.
  • TripNet outperforms prior models on drag, pressure, and wall-shear prediction while reducing CFD runtime from 10 hours to seconds.It reports R2 scores of 0.972 on DrivAerNet and 0.957 on DrivAerNet++ for drag, with leading surface-field errors across the evaluated datasets.
  • A unified triplane backbone supports drag coefficients, surface pressure and wall shear stress, and full 3D flow-field prediction.Different task-specific models can operate on the shared representation, including a U-Net plus MLP for fields and a lightweight CNN for drag.
  • The authors present TripNet as, to their knowledge, the first use of triplanes for solving PDEs rather than exclusively for generative modeling.

2 Method

TripNet fits each car’s geometry into three orthogonal feature planes and decodes queried features into aerodynamic quantities. Task-specific CNN or U-Net–MLP architectures use this compact representation to predict scalar, surface, and volumetric fields.

  • 2.1 Triplane Representation: TripNet converts each car into three orthogonal triplanes whose projected features are aggregated by an MLP to reconstruct occupancy.The representation uses Pxy, Pyz, and Pxz planes and maps 3D coordinates to an occupancy prediction.
  • 2.1 Triplane Representation: Two-stage fitting jointly trains triplanes and an MLP decoder on 500 objects, then freezes the decoder while fitting each car’s triplanes individually.The fitting objective includes regularization intended to reduce compression artifacts and smooth or constrain feature planes.
  • 2.1 Triplane Representation: Triplanes provide an O(n^2) geometric representation that preserves spatial information more efficiently than O(n^3) 3D geometry models.
  • Drag coefficient prediction: A lightweight CNN predicts drag from channel-pooled triplanes, compressing each plane’s 32 channels with max, min, and mean pooling.
  • Surface and volume field prediction: A hybrid U-Net–MLP predicts pressure, wall shear stress, and velocity fields by combining global multi-scale features with localized queried-feature refinement.The U-Net processes concatenated triplane features, while the MLP maps queried features to local aerodynamic quantities.
  • Evaluation: Evaluation uses DrivAerNet as a constrained single-category benchmark and DrivAerNet++ as a broader multi-category generalization task.

3 Results

TripNet achieves strong accuracy across drag, surface-field, and volumetric flow prediction while scaling to large aerodynamic simulations. Its triplane representation preserves geometric information, supports efficient query-based prediction, and extends across multiple CFD tasks.

  • 3.1 Aerodynamic Drag Prediction: TripNet achieves the highest drag-prediction accuracy on both DrivAerNet and DrivAerNet++, with R2 scores of 0.972 and 0.957, respectively.The comparison covers point-cloud, graph, and other deep-learning baselines on industry-standard car datasets.
  • 3 Results: Triplane features preserve full geometric information implicitly, enabling precise aerodynamic predictions without the downsampling losses of mesh, voxel, and point-cloud methods.For surface fields, a U-Net refines spatial coherence and an MLP enables query-based predictions that reduce noise and improve accuracy.
  • 3.2 Surface Field: TripNet matches or exceeds state-of-the-art performance for surface pressure and wall shear stress prediction across both DrivAerNet datasets.It attains the lowest or second-lowest values across MSE, MAE, Max AE, ℓ1, and ℓ2 error metrics.
  • 3.3 Volume Flow Field: TripNet’s volumetric flow predictions achieve 7.15% ℓ1 and 10.7% ℓ2 error for Ux on the DrivAerNet++ test set.Uy and Uz have higher errors because they are sparser and lower-magnitude components.
  • 3.3 Volume Flow Field: TripNet accurately reproduces wake flow behind the car and the wake generated by its side mirrors across three cross-sectional planes.The visualization compares predictions with CFD on the symmetry plane, wake region, and a vertical slice through the car body.
  • 3 Results: TripNet’s scalability is evaluated through relative L2 error, GPU memory usage, and inference time across increasing point resolutions.These comparisons directly assess how the method behaves as prediction resolution grows.

4 Conclusion

TripNets provide compact triplane-based surrogate models for drag, surface-field, and volumetric-flow prediction, achieving state-of-the-art aerodynamic performance. The representation also captures localized flow characteristics and supports efficient modeling across tasks.

  • TripNets support drag coefficient, surface pressure, wall shear stress, and 3D flow-field prediction through compact triplane features.The framework functions as a surrogate model across multiple aerodynamic tasks.
  • For drag prediction on DrivAerNet, the model reaches an R2 score of 0.97 with one hour of training and 0.01-second inference per car.
  • TripNet achieves state-of-the-art pressure and wall shear stress prediction across DrivAerNet and DrivAerNet++, while using 4.1× less GPU memory and 1.4× faster inference than FigConvNet at the highest resolution.Its nearly constant memory usage and inference time arise from the resolution-independent triplane representation.
  • Triplane features encode geometric information and localized flow characteristics, enabling accurate volumetric flow predictions near the car surface.

5 Limitation and Future Work

The method’s generalization beyond its training setting remains unverified, and it does not explicitly enforce physical constraints. Future work therefore targets broader geometries, flow conditions, and physics-informed modeling.

  • Generalization to topologically distinct shapes such as F1 cars and to flow regimes beyond the training Reynolds number remains unverified.Triplane fitting is performed separately for each geometry.
  • Without explicit mass-conservation or turbulence-consistency constraints, the model may produce non-physical artifacts in regions with sparse training data.

Technical Appendices and Supplementary Material

The supplementary material situates TripNet among neural simulation models and details its training setup, triplane fitting procedure, scalability, and query-based implicit representation. Its central distinction is resolution-independent prediction beyond input points.

  • Deep Learning Models for Simulation: Scaling transformer-based solvers to industrial datasets such as DrivAerNet++, with millions of nodes, remains challenging.
  • Related Work: TripNet uses a triplane-based implicit representation that supports query-based predictions beyond input points while maintaining memory efficiency across resolutions.
  • Implicit and Neural Representations: Implicit representations enable solution queries at arbitrary spatial locations, unlike graph and point-cloud models that predict only at discrete mesh nodes.
  • Training Setup: The supplementary training setup uses 4 H100 GPUs for TripNet, 200 epochs, batch size 32, and 200,000 sampled surface points, with a one-A100 comparison setup.
  • Triplane Fitting: Triplane fitting uses two stages: joint training on 500 cars followed by individual fitting with a shared MLP decoder for DrivAerNet++.

B.1.2 Triplane Resolution

Higher triplane resolution improves geometry reconstruction fidelity but increases computational cost, motivating a resolution choice that balances accuracy and efficiency. The representation supports fixed-dimensional, query-based scaling across simulation resolutions.

  • Triplane Resolution: Increasing triplane resolution improves reconstruction accuracy, precision, recall, and F1 score, while lower resolution loses fine surface details.
  • Triplane Resolution: At 32 × 256 × 256, reconstructed geometry closely matches the original mesh and reaches an F1-score of 0.9936.
  • Triplane Resolution: A 32 × 128 × 128 resolution offers a balance between reconstruction accuracy and computational efficiency for downstream tasks.
  • Scalability: TripNet scales through query-based prediction over fixed-dimensional triplane feature maps rather than complete graphs limited by GPU memory.

C.1 Drag coefficient

TripNet uses a lightweight CNN over stacked triplanes to predict drag coefficients, while related visual evaluations compare aerodynamic field predictions and errors across unseen car designs.

  • C.1 Drag coefficient: A lightweight CNN predicts the drag coefficient from stacked 9-channel triplanes using an encoder-style architecture followed by fully connected regression layers.The input tensor has shape 9 × 128^2 and contains triplanes Pxy, Pyz, and Pxz.
  • C.1 Drag coefficient: The drag-coefficient evaluation is presented on the DrivAerNet++ test set in Table 5.
  • C.1 Drag coefficient: The supplied figures compare model predictions and absolute error distributions for pressure and wall shear stress across three unseen car designs against ground-truth CFD.The comparisons include multiple car categories and underbody and wheel configurations to assess generalization.

C.4 Volumetric Flow Field

TripNet predicts volumetric aerodynamic fields with triplane-based representations and is evaluated against CFD across multiple cross-sections. The surrounding formulation uses turbulence-modeling and drag-related fluid quantities as physical context.

  • C.4 Volumetric Flow Field: TripNet predictions closely match CFD across upstream and downstream cross-sections, capturing key flow features near the vehicle surface and in the wake.Figure 8 compares volumetric field predictions from CFD and TripNet at multiple locations along the car body.
  • C.4 Volumetric Flow Field: The Navier–Stokes framework models three-dimensional fluid behavior using velocity, pressure, density, viscosity, and body-force terms under incompressible constant-density and constant-viscosity assumptions.The equations provide the physical basis for the simulated flow quantities.
  • C.4 Volumetric Flow Field: The k−ω SST model supplies the RANS turbulence framework used to generate the DrivAerNet and DrivAerNet++ simulations.The model solves for turbulence kinetic energy k and specific dissipation rate ω and is effective for flow separation.
  • C.4 Volumetric Flow Field: The turbulence equations describe transport of k and ω, with diffusivity, production, model-constant, and source terms representing turbulence effects.The formulation includes Dk, Dω, G, γ, β, β*, Sk, and Sω as described in the paper.
  • C.4 Volumetric Flow Field: The paper defines total drag as the sum of pressure-drag and wall-shear-stress contributions, with force normalization using dynamic pressure and reference area.Pressure contribution depends on surface normals, cell areas, pressure, and motion direction; shear contribution depends on wall shear stress and cell areas.
  • C.4 Volumetric Flow Field: Surface-field prediction can either support drag estimation by integrating predicted pressure and wall shear stress or provide interpretable regional aerodynamic-force information.

E Description of the DrivAerNet++ Dataset

DrivAerNet++ is a large multimodal dataset of industry-standard cars with high-fidelity CFD, geometric, aerodynamic, and flow-field information across varied vehicle configurations.

  • E Description of the DrivAerNet++ Dataset: DrivAerNet++ contains over 8,000 unique industry-standard car models generated with high-fidelity CFD simulations.Its body styles include fastback, notchback, and estateback designs, with varied underbody and wheel configurations.
  • E Description of the DrivAerNet++ Dataset: Each model includes 3D meshes, parametric models, aerodynamic coefficients, flow and surface fields, segmented parts, and point-cloud data.These modalities support design optimization, generative modeling, surrogate training, CFD acceleration, and geometric classification.
  • E Description of the DrivAerNet++ Dataset: The dataset compares drag-coefficient ranges, means, and sample counts across car configurations normalized by projected frontal area.Configurations vary by smooth or detailed underbody and wheel/tire setup.

F Evaluation Metrics

The evaluation uses task-specific metrics: relative L1 and L2 errors for surface fields, R2 for drag coefficients, and MSE for volumetric flow fields.

  • F Evaluation Metrics: MAE summarizes average absolute prediction error, while Max AE reports the largest absolute error as a worst-case measure.The maximum absolute error identifies the greatest absolute difference between prediction and ground truth.
  • F Evaluation Metrics: Surface pressure and wall shear stress predictions are evaluated using relative L1 and relative L2 errors.These metrics normalize prediction differences relative to the magnitude of the ground-truth fields.
  • F Evaluation Metrics: Drag coefficient estimation is ranked using R2, which measures how well predictions explain variance in the ground-truth values.The mean ground-truth value is used in the R2 definition.
  • F Evaluation Metrics: Volumetric flow-field prediction is assessed with mean squared error, which averages squared prediction differences and penalizes larger errors more heavily.
  • F Evaluation Metrics: The metric set jointly evaluates accuracy, robustness, and worst-case performance across scalar, surface-field, and volumetric tasks.
Loading 2503.17400v2…