TripOptimizer: Generative 3D Shape Optimization and Drag Prediction using Triplane VAE Networks

The method

Optimize the shape by descending through the surrogate

TripOptimizer is a fully differentiable deep-learning framework for rapid aerodynamic analysis and shape optimization. A triplane variational autoencoder compresses 3D car geometry into a structured latent space; a drag-prediction head evaluates aerodynamics directly from that latent. Because the whole pipeline is differentiable, drag gradients flow back into the latent space — turning shape optimization into gradient descent.

Does descending through a neural network actually produce better cars? In case studies on industry-standard geometries, the optimizer delivered drag-coefficient reductions of up to 11.8% — and the claim wasn't left to the surrogate to grade its own homework: the optimized shapes were re-simulated with independent high-fidelity CFD using more than 150 million cells. A quiet superpower of the implicit triplane representation: it shrugs off geometric imperfections, so even non-watertight meshes — the kind that break traditional adjoint-based optimization outright — can be optimized directly. That makes the framework practical in early design stages, where geometry is messy and CFD budgets are thin.

Generative shape optimization: geometry evolves along the drag gradient in the triplane latent space.

The TripOptimizer pipeline: a point cloud is encoded to a latent vector, navigated toward lower drag via the prediction head's gradient, decoded through triplanes and an occupancy generator into an optimized geometry, then CFD-validated. — The fully differentiable loop: encode geometry → descend the drag gradient in latent space → decode triplanes into an optimized shape → validate with CFD.

Contributions

What the paper delivers

Triplane geometry encoding

Three axis-aligned feature planes represent full 3D car shapes compactly, capturing fine geometric detail without volumetric memory costs.

Drag prediction from latents

Aerodynamic performance is predicted directly in latent space — no meshing, no simulation in the loop.

Gradient-based 3D design

End-to-end differentiability enables generative optimization of industry-standard designs, producing lower-drag variants in minutes rather than CFD-weeks.

Results

Accurate drag, controllable shape

The latent-space drag head reaches R² = 0.93 on unseen industry-standard designs with consistent error across fastback, estateback, and notchback categories — and the latent space supports controlled, localized shape edits rather than arbitrary deformations.

Predicted versus true drag coefficient scatter plot with R-squared of 0.93, and absolute error box plots per vehicle category. — Drag prediction on unseen designs: R² = 0.93, with comparable error across all three vehicle categories.

Reference

Citation

@article{vatani2025tripoptimizer,
  title   = {TripOptimizer: Generative Three-Dimensional Shape Optimization
             and Drag Prediction using Triplane VAE Networks},
  author  = {Vatani, Parsa and Elrefaie, Mohamed and Nazarpour, Farhad
             and Ahmed, Faez},
  journal = {Physics of Fluids},
  volume  = {37},
  number  = {12},
  pages   = {127113},
  year    = {2025}
}

Keep exploring