Research map / Probabilistic and scientific modeling / Molecules and proteins
Aerodynamic Coefficients Prediction via Cross-Attention Fusion and Physical-Informed Training
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 a cross-attention transformer that fuses PointGPT-pretrained shape features with flow-condition encodings and adds physics-based consistency terms to predict aerodynamic coefficients.
Problem
Data-driven prediction of aerodynamic coefficients from 3D shape and flow conditions suffers from limited high-quality aerodynamic data and poor generalization to new shapes and flow conditions.
Approach
Separates shape encoding (PointGPT-pretrained point cloud transformer) from flow-condition encoding (MLP), fuses them with a cross-attention decoder, and adds plug-and-play physics equations (AoA-based transforms of CD/CL into Cx, Cz, and pitch moment) as extra loss terms computed from predicted outputs.
Key results
- On DrivAerNet++ drag prediction, the 100c64n shape+LWH model reaches MSE 5.275e-5, R2 0.962, versus PointNet at 14.9e-5 and R2 0.643.
- On AircraftData, cross-attention fusion achieves total MRE 6.728e-2 (MAE 13.304e-3 etc.), better than concatenation (MRE total 7.185) and multiplication (7.746) in MAE terms.
- Physical-informed fine-tuning gives the lowest total MRE of 4.610e-2 in the training-strategy ablation, versus 5.504e-2 for the base model.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add vehicle length, width, and height as flow features (ablation) | Proposed model with 100c64n · DrivAerNet++ | MSE (×10−5): 5.293 -> 5.275 | Helped |
| Increase point-cloud sampling from 64c32n to 100c64n (ablation) | Proposed shape-only model · DrivAerNet++ | MSE (×10−5): 9.031 -> 5.293 | Helped |
| Use cross-attention instead of switching attention (ablation) | AircraftData fusion model · AircraftData | MRE (×10−2), Total: 6.996 -> 6.728 | Helped |
| Use cross-attention instead of feature concatenation (ablation) | AircraftData fusion model · AircraftData | MRE (×10−2), Total: 7.185 -> 6.728 | Helped |
| Use cross-attention instead of feature multiplication (ablation) | AircraftData fusion model · AircraftData | MRE (×10−2), Total: 7.746 -> 6.728 | Helped |
| Add shape feature reuse (ablation) | AircraftData training strategy · AircraftData | MRE (×10−2), Total: 4.793 -> 4.760 | Helped |
| Fine-tune the pre-trained model (ablation) | AircraftData training strategy · AircraftData | MRE (×10−2), Total: 4.793 -> 4.582 | Helped |
| Add physical-informed training (ablation) | AircraftData training strategy · AircraftData | MRE (×10−2), Total: 4.793 -> 4.676 | Helped |
| Combine physical-informed training with fine-tuning (ablation) | AircraftData training strategy · AircraftData | MRE (×10−2), Total: 4.793 -> 4.610 | Helped |
Surprising findings
- Multiplication fusion yields the best MREn (4.263e-2) in Table 2 despite the worst MRE, and the paper does not clearly explain this discrepancy.
Limitations
- Automotive benchmark uses a single flow condition, so cross-attention's advantage cannot be fully exercised there.
- The aircraft dataset is self-generated and not released, limiting reproducibility.
- Physics-informed gains are modest (e.g., total MRE 4.610 vs 4.582 for plain fine-tuning in some cases) and the 'moderate' use of physics is not quantified.
Experimental setup
Datasets: DrivAerNet++ AircraftData
Benchmarks: DrivAerNet++ AircraftData
Metrics: MSE MAE MaxAE R2 MRE MREn
Baselines: PointNet GCNN RegDGCNN Concatenation fusion Multiplication fusion Switch attention
Builds on: PointGPT PINN DrivAerNet++
Closest papers by meaning
- AeroGTO: An Efficient Graph-Transformer Operator for Learning Large-Scale Aerodynamics of 3D Vehicle Geometries (AAAI 2025 · similarity 0.86)
- DrivAerStar: An Industrial-Grade CFD Dataset for Vehicle Aerodynamic Optimization (NeurIPS 2025 · similarity 0.84)
- WindsorML: High-Fidelity Computational Fluid Dynamics Dataset For Automotive Aerodynamics (NeurIPS 2024 · similarity 0.84)
- 3DID: Direct 3D Inverse Design for Aerodynamics with Physics-Aware Optimization (NeurIPS 2025 · similarity 0.84)
- AFBench: A Large-scale Benchmark for Airfoil Design (NeurIPS 2024 · similarity 0.83)
- DrivAerNet++: A Large-Scale Multimodal Car Dataset with Computational Fluid Dynamics Simulations and Deep Learning Benchmarks (NeurIPS 2024 · 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.