Research map / 3D vision / Face and shape
3DID: Direct 3D Inverse Design for Aerodynamics with Physics-Aware Optimization
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 3DID, a 3D inverse-design framework that learns a unified physics-geometry triplane latent (PG-VAE), then runs objective-guided latent diffusion followed by FFD-based topology-preserving refinement, evaluated on DrivAerNet++ aerodynamic car drag optimization.
Problem
Inverse design of 3D geometries under physics objectives is hard because the 3D design space is huge and prior methods either use 2D proxies or refine an initial shape, limiting exploration and validity.
Approach
Couples a continuous physics-geometry triplane latent with a two-stage pipeline: gradient-guided latent diffusion from noise for global exploration, then gradient-based free-form deformation refinement against a GNN surrogate that preserves mesh topology.
Key results
- 3DID reduces simulated drag (Sim-Drag) to 0.3536 versus 0.4066 for the best baseline (TripNet), a 13% relative reduction as reported.
- 3DID achieves the highest novelty (1.1709) among all compared methods, while its coverage drops to 0.4300, lower than most baselines.
- Ablation: 3DID outperforms CEM, GD and unguided diffusion on Pred-Drag, Sim-Drag and Novelty, but unguided diffusion has the highest coverage (0.7104).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the full 3DID pipeline rather than the strongest baseline for simulated drag | 3DID on DrivAerNet++ vehicle aerodynamic design · DrivAerNet++ vehicle aerodynamic design | Sim-Drag: 0.4097 -> 0.3536 | Helped |
| use the full 3DID pipeline rather than the strongest baseline for predicted drag | 3DID on DrivAerNet++ vehicle aerodynamic design · DrivAerNet++ vehicle aerodynamic design | Pred-Drag: 0.2951 -> 0.2607 | Helped |
| use the full 3DID pipeline rather than the strongest baseline for novelty | 3DID on DrivAerNet++ vehicle aerodynamic design · DrivAerNet++ vehicle aerodynamic design | Novelty: 1.0399 -> 1.1709 | Helped |
| use topology-preserving refinement (ablation) | 3DID on DrivAerNet++ vehicle aerodynamic design · DrivAerNet++ vehicle aerodynamic design | Sim-Drag: 0.3766 -> 0.3536 | Helped |
| use topology-preserving refinement (ablation) | 3DID on DrivAerNet++ vehicle aerodynamic design · DrivAerNet++ vehicle aerodynamic design | Coverage: 0.6950 -> 0.4300 | Hurt |
| use unified physics–geometry representation rather than geometry-only TripNet (ablation) | 3DID on DrivAerNet++ vehicle aerodynamic design · DrivAerNet++ vehicle aerodynamic design | Sim-Drag: 0.4066 -> 0.3536 | Helped |
| use unified physics–geometry representation rather than geometry-only TripNet (ablation) | 3DID on DrivAerNet++ vehicle aerodynamic design · DrivAerNet++ vehicle aerodynamic design | Novelty: 1.0683 -> 1.1709 | Helped |
| use objective-gradient guidance during diffusion sampling (ablation) | 3DID on DrivAerNet++ vehicle aerodynamic design · DrivAerNet++ vehicle aerodynamic design | Sim-Drag: 0.3944 -> 0.3536 | Helped |
| remove objective-gradient guidance from diffusion sampling (ablation) | 3DID on DrivAerNet++ vehicle aerodynamic design · DrivAerNet++ vehicle aerodynamic design | Coverage: 0.4300 -> 0.7104 | Helped |
| replace the two-stage optimization pipeline with CEM (ablation) | 3DID on DrivAerNet++ vehicle aerodynamic design · DrivAerNet++ vehicle aerodynamic design | Sim-Drag: 0.3536 -> 0.3987 | Hurt |
| replace the two-stage optimization pipeline with gradient descent (ablation) | 3DID on DrivAerNet++ vehicle aerodynamic design · DrivAerNet++ vehicle aerodynamic design | Sim-Drag: 0.3536 -> 0.4095 | Hurt |
Surprising findings
- Topology-preserving refinement improves drag but lowers coverage (0.6950 without refinement to 0.4300 with it), a trade-off the authors attribute to designs moving beyond the training distribution.
- Unguided diffusion attains the highest coverage, indicating it mimics the dataset distribution rather than discovering better designs.
Limitations
- Restricted to static physical fields; time-dependent problems are not supported.
- Only single-objective optimization is handled.
- Physics awareness comes from data-driven surrogates and does not enforce hard physical constraints.
Experimental setup
Datasets: DrivAerNet++
Benchmarks: DrivAerNet++
Metrics: Pred-Drag Sim-Drag Novelty Coverage
Baselines: Cross-Entropy Method Gaussian-process Bayesian optimization Backprop Voxel VAE Voxel+PCA TripNet W/O Guidance
Builds on: DiT MeshGraphNet DrivAerNet++ free-form deformation triplane representation
Closest papers by meaning
- Compositional Generative Inverse Design (ICLR 2024 · similarity 0.89)
- AFBench: A Large-scale Benchmark for Airfoil Design (NeurIPS 2024 · similarity 0.88)
- EngiBench: A Framework for Data-Driven Engineering Design Research (NeurIPS 2025 · similarity 0.85)
- Co-PiLOT: Constrained Physics-Informed Latent Optimization for Target-Driven Inverse Design (NeurIPS 2026 · similarity 0.85)
- AeroGTO: An Efficient Graph-Transformer Operator for Learning Large-Scale Aerodynamics of 3D Vehicle Geometries (AAAI 2025 · similarity 0.85)
- GAIA: Geometry-Adaptive Operator Learning for Forward and Inverse Problems (NeurIPS 2026 · similarity 0.85)
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.