Research map / Probabilistic and scientific modeling / Molecules and proteins
DragSolver: A Multi-Scale Transformer for Real-World Automotive Drag Coefficient Estimation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DragSolver, a multi-scale Transformer with heterogeneous scale normalization, surface-guided gating, and MC dropout uncertainty to predict automotive drag coefficient from 3D meshes/point clouds.
Problem
Rapid and accurate drag coefficient (Cd) estimation from large-scale, heterogeneous real-world 3D vehicle geometries, which CFD and wind tunnel tests are too slow to support in iterative design.
Approach
Combines multi-stage Transformer encoder with multi-scale global pooling, wheelbase-based and hybrid-sampling normalization, multi-stage gating to suppress interior geometry, and Monte Carlo dropout for epistemic uncertainty in a geometry-to-Cd regression.
Key results
- In-distribution on DrivAerNet++, DragSolver reaches relative L2 0.0005 and R2 0.9712, a 64.2% error promotion over the second-best baseline.
- Cross-dataset (DrivAerNet++ to DrivAerML) relative L2 0.0091 with R2 0.2732, while Mamba3D gets R2 -0.0684.
- Under interior mismatch, gating reduces relative L2 from 0.0057 to 0.0024 (ablation); inference ~0.93 s per sample, ~4.8 s with MC dropout.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DragSolver instead of the strongest baseline, Mamba3D | Automotive drag coefficient prediction · DrivAerNet | Relative L2: 0.0015 -> 0.0007± 0.0001 | Helped |
| Use DragSolver instead of the strongest baseline, Mamba3D | Automotive drag coefficient prediction · DrivAerNet++ | Relative L2: 0.0014 -> 0.0005± 0.0001 | Helped |
| Use DragSolver instead of Mamba3D | Cross-shape prediction on DrivAerNet++ estateback (train F+N) · DrivAerNet++ estateback | Relative L2: 0.0097 -> 0.0031± 0.0003 | Helped |
| Use DragSolver instead of Mamba3D | Cross-shape prediction on DrivAerNet++ notchback (train F+E) · DrivAerNet++ notchback | Relative L2: 0.0054 -> 0.0034± 0.0003 | Helped |
| Use DragSolver instead of Mamba3D | Cross-dataset prediction · DrivaerML (train DrivAerNet++) | Relative L2: 0.0197 -> 0.0091± 0.0001 | Helped |
| Add multi-scale feature fusion (ablation) | DragSolver · DrivAerNet++ in-distribution | Relative L2: 0.0049 -> 0.0026 | Helped |
| Add multi-scale feature fusion (ablation) | DragSolver · DrivAerNet++ cross-shape | Relative L2: 0.0081 -> 0.0065 | Helped |
| Add heterogeneous scale normalization (ablation) | DragSolver with multi-scale features · DrivAerNet++ in-distribution | Relative L2: 0.0026 -> 0.0007 | Helped |
| Add surface-guided gating (ablation) | DragSolver with multi-scale features and normalization · DrivAerNet++ interior mismatch | Relative L2: 0.0057 -> 0.0024 | Helped |
| Set multi-scale depth to L=5 rather than L=1 (ablation) | DragSolver · DrivAerNet++ | Relative L2: 0.0020 -> 0.0008 | Helped |
| Use max instead of mean global pooling (ablation) | DragSolver · DrivAerNet++ | Relative L2: 0.0009 -> 0.0007 | Helped |
| Increase voxel size from 0.01 to 0.05 or 0.1 (ablation) | DragSolver · DrivAerNet++ | Relative L2: 0.0009 -> 0.0016; 0.0009 -> 0.0029 | Hurt |
Surprising findings
- Several neural operator baselines (Transolver, GINO, MeshGraphNet) give negative R2 on the automotive datasets.
- Cross-dataset R2 for DragSolver is only 0.2732, indicating limited transfer despite low relative L2.
Limitations
- Cross-dataset R2 remains low (0.2732), so generalization to different datasets is limited.
- Uncertainty calibration is not evaluated beyond reporting standard deviations; future work mentions expanded calibration.
- Code is promised only upon acceptance; no release link is given in the text.
Experimental setup
Datasets: DrivAerNet++ DrivAerNet DrivAerML
Benchmarks: DrivAerNet DrivAerNet++ DrivAerML
Metrics: relative L2 error R2 score
Baselines: PointNet PointNet++ PointTransformer PointGPT Mamba3D PointTransformer V3 MeshGraphNet RegDGCNN GINO Transolver GNOT
Builds on: Point Transformer V3 Monte Carlo dropout Transolver Point-BERT
Closest papers by meaning
- AeroGTO: An Efficient Graph-Transformer Operator for Learning Large-Scale Aerodynamics of 3D Vehicle Geometries (AAAI 2025 · similarity 0.89)
- DrivAerStar: An Industrial-Grade CFD Dataset for Vehicle Aerodynamic Optimization (NeurIPS 2025 · similarity 0.88)
- DrivAerNet++: A Large-Scale Multimodal Car Dataset with Computational Fluid Dynamics Simulations and Deep Learning Benchmarks (NeurIPS 2024 · similarity 0.87)
- AutoHood3D: A Multi‑Modal Benchmark for Automotive Hood Design and Fluid–Structure Interaction (NeurIPS 2025 · similarity 0.85)
- WindsorML: High-Fidelity Computational Fluid Dynamics Dataset For Automotive Aerodynamics (NeurIPS 2024 · similarity 0.84)
- Aerodynamic Coefficients Prediction via Cross-Attention Fusion and Physical-Informed Training (AAAI 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.