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

TL;DR

TripOptimizer addresses the high computational cost of traditional aerodynamic shape optimization by delivering a fully differentiable, point-cloud–based framework that jointly predicts drag and reconstructs 3D geometry via a triplane implicit representation. It introduces encoder-parameter refinement to efficiently steer designs toward target $C_d$ without destroying geometric plausibility, trained on the DrivAerNet++ dataset to achieve $R^2=0.93$ for $C_d$ and high geometric fidelity. The resulting designs achieve CFD-validated drag reductions up to $11.8\%$, while remaining robust to non-watertight meshes, enabling rapid exploration in early-stage automotive design. The learned latent space organizes shapes by topology and aerodynamics, facilitating efficient design-space navigation and offering actionable outputs like signed-distance maps for CAD refinement.

Abstract

The computational cost of traditional Computational Fluid Dynamics-based Aerodynamic Shape Optimization severely restricts design space exploration. This paper introduces TripOptimizer, a fully differentiable deep learning framework for rapid aerodynamic analysis and shape optimization directly from vehicle point cloud data. TripOptimizer employs a Variational Autoencoder featuring a triplane-based implicit neural representation for high-fidelity 3D geometry reconstruction and a drag coefficient prediction head. Trained on DrivAerNet++, a large-scale dataset of 8,000 unique vehicle geometries with corresponding drag coefficients computed via Reynolds-Averaged Navier-Stokes simulations, the model learns a latent representation that encodes aerodynamically salient geometric features. We propose an optimization strategy that modifies a subset of the encoder parameters to steer an initial geometry towards a target drag value, and demonstrate its efficacy in case studies where optimized designs achieved drag coefficient reductions up to 11.8\%. These results were subsequently validated by using independent, high-fidelity Computational Fluid Dynamics simulations with more than 150 million cells. A key advantage of the implicit representation is its inherent robustness to geometric imperfections, enabling optimization of non-watertight meshes, a significant challenge for traditional adjoint-based methods. The framework enables a more agile Aerodynamic Shape Optimization workflow, reducing reliance on computationally intensive CFD simulations, especially during early design stages.

Figures (18)

• Figure 1: Overview of our proposed TripOptimizer, a fully differentiable generative triplane-based model for aerodynamic analysis and shape optimization. The model jointly reconstructs high-fidelity 3D vehicle shapes and predicts their drag coefficients ($C_d$), then optimizes designs towards user-defined aerodynamic targets.
• Figure 2: A selection of diverse vehicle shapes within the DrivAerNet++ dataset Elrefaie2024DrivAerNetPP. The figure covers Estatebacks, Notchbacks, and Fastbacks.
• Figure 3: The data transformation visualization, with the original geometry (left), corresponding surface point cloud (middle), and three orthogonal slices of semi-continuous occupancy field (right).
• Figure 4:
• Figure 5: Visualization of the Triplane-based Occupancy Decoder components and example triplane of geometry ID F-S-WWC-WM-260 from the DrivAerNet++ dataset.