TripNet: Learning Large-scale High-fidelity 3D Car Aerodynamics with Triplane Networks

TL;DR

TripNet introduces a triplane-based implicit 3D geometry representation to build a scalable, high-fidelity CFD surrogate capable of querying predictions at arbitrary locations. A shared triplane backbone with task-specific heads enables rapid prediction of drag, surface fields, and full 3D flow, achieving state-of-the-art accuracy on DrivAerNet and DrivAerNet++ while dramatically reducing memory and computation compared with mesh- and graph-based methods. The approach demonstrates strong performance across multiple aerodynamic tasks, including near-surface fields and volumetric flow, with favorable inference times and scalability to large cell counts. Limitations include geometry-specific fitting and lack of explicit physics constraints, guiding future work toward physics-informed training and broader geometric generalization. Overall, TripNet offers a practical, memory-efficient surrogate for industrial CFD workflows, enabling fast design iteration and analysis on high-resolution geometries.

Abstract

Surrogate modeling has emerged as a powerful tool to accelerate Computational Fluid Dynamics (CFD) simulations. Existing 3D geometric learning models based on point clouds, voxels, meshes, or graphs depend on explicit geometric representations that are memory-intensive and resolution-limited. For large-scale simulations with millions of nodes and cells, existing models require aggressive downsampling due to their dependence on mesh resolution, resulting in degraded accuracy. We present TripNet, a triplane-based neural framework that implicitly encodes 3D geometry into a compact, continuous feature map with fixed dimension. Unlike mesh-dependent approaches, TripNet scales to high-resolution simulations without increasing memory cost, and enables CFD predictions at arbitrary spatial locations in a query-based fashion, independent of mesh connectivity or predefined nodes. TripNet achieves state-of-the-art performance on the DrivAerNet and DrivAerNet++ datasets, accurately predicting drag coefficients, surface pressure, and full 3D flow fields. With a unified triplane backbone supporting multiple simulation tasks, TripNet offers a scalable, accurate, and efficient alternative to traditional CFD solvers and existing surrogate models.

Figures (15)

• Figure 1: Illustration of Triplane Networks and their role in CFD tasks. The figure demonstrates how triplane features encode spatial information for aerodynamic analysis, showcasing their applications in predicting surface pressure, wall shear stress, and full 3D flow fields.
• Figure 1: Visulization of the raw triplane features of F_D_WM_WW_1282 (fastback with detailed underbody, with wheels, and with mirrors) used for surface field and 3D flow field predictions.
• Figure 2: Comparison of pressure predictions and errors for car design E_S_WWC_WM_094 (estateback with smooth underbody, wheels closed, and with mirrors) from the unseen test set. The first row shows the input mesh followed by predictions from RegDGCNN, FigConvNet, Transolver, TripNet (ours), and the ground-truth CFD. The second row highlights the absolute difference between the predictions and ground-truth CFD, visualizing the absolute error distribution for each model.
• Figure 2: 3D Geometry reconstructed by triplanes with different dimensions.
• Figure 3: Comparison of 3D flow field predictions, including velocity magnitude ($U$) and errors, for the car design E_S_WWC_WM_094 (estateback with smooth underbody, with wheels closed, and with mirrors) from the unseen test set. The visualization includes three planes: $y=0$ (symmetry plane), $z=1$, and $x=4$ (wake of the car). The first row shows the ground truth CFD results, the second row displays our model's predictions, and the final row illustrates the absolute error.