TripNet: Learning large-scale high-fidelity three-dimensional car aerodynamics with triplane networks
- 1Department of Mechanical Engineering, Massachusetts Institute of Technology 1 , Cambridge, Massachusetts 02139,
- 2Technical University of Munich 2 , Garching,
Abstract
Surrogate modeling has become a powerful approach for accelerating computational fluid dynamics (CFD) simulations, yet existing three-dimensional (3D) geometric learning models–based on meshes, point clouds, or voxels–remain fundamentally limited by resolution, memory cost, and dependence on explicit geometric discretizations. We introduce TripNet, a triplane-based neural framework that encodes 3D geometry into a compact, continuous, resolution-independent representation. Unlike traditional mesh-dependent models, TripNet supports query-based predictions at arbitrary spatial locations, enabling high-fidelity aerodynamic field estimation without relying on mesh connectivity or downsampled geometry. TripNet achieves state-of-the-art performance on the DrivAerNet and DrivAerNet++ datasets across three major aerodynamic tasks: drag coefficient prediction, surface pressure and wall-shear stress regression, and full 3D volumetric flow-field prediction. TripNet attains an R2 of 0.972 for drag prediction and a relative L2 error of 10.39% for velocity magnitude. TripNet reduces inference time by several orders of magnitude–predicting drag in 0.01 s and full 3D flow fields in 2 s on a single graphics processing unit–while improving accuracy and reducing memory cost compared to strong baselines such as FigConvNet, Transolver, and MeshGraphNet. Beyond in-distribution prediction, we demonstrate that triplane features encode robust, transferable geometric–aerodynamic priors. TripNet exhibits strong cross-shape generalization, outperforming prior works when trained on two car-body classes and tested on an unseen third shape. Furthermore, TripNet achieves data-efficient cross-dataset transfer: when pretrained on DrivAerNet++ and finetuned on the high-fidelity DrivAerML dataset, TripNet surpasses models trained solely on DrivAerML across all metrics, even with as few as 10–50 gradient steps. Collectively, these results highlight triplanes as a scalable, generalizable representation for partial differential equation surrogate modeling and establish TripNet as an accurate, efficient alternative to traditional CFD solvers and existing geometric deep learning models.
Related
- HiLiftAeroML: A High-Fidelity Computational Fluid Dynamics Dataset for High-Lift Aircraft Aerodynamics
- SHIFT-Truck: A High-Fidelity Aerodynamics Dataset and Benchmark for Pickup Trucks
- High-fidelity quantum gate learning via physics-informed neural networks
- High-fidelity state transfer in SiV-phononic networks with closed mechanical subsystems
- High-fidelity Three-Dimensional Flow Field Reconstruction Driven by Turbulent Fluid Dynamics
BibTeX
@article{chen2026tripnet,
title = {{TripNet: Learning large-scale high-fidelity three-dimensional car aerodynamics with triplane networks}},
author = {Q Chen and Mohamed Elrefaie and Angela Dai and Faez Ahmed},
journal = {Physics of Fluids},
year = {2026},
month = jun,
doi = {10.1063/5.0324695},
url = {https://doi.org/10.1063/5.0324695},
note = {CC BY},
}
RIS
TY - JOUR
TI - TripNet: Learning large-scale high-fidelity three-dimensional car aerodynamics with triplane networks
AU - Q Chen
AU - Mohamed Elrefaie
AU - Angela Dai
AU - Faez Ahmed
PY - 2026
DA - 2026/06/01
JO - Physics of Fluids
DO - 10.1063/5.0324695
UR - https://doi.org/10.1063/5.0324695
AB - Surrogate modeling has become a powerful approach for accelerating computational fluid dynamics (CFD) simulations, yet existing three-dimensional (3D) geometric learning models–based on meshes, point clouds, or voxels–remain fundamentally limited by resolution, memory cost, and dependence on explicit geometric discretizations. We introduce TripNet, a triplane-based neural framework that encodes 3D geometry into a compact, continuous, resolution-independent representation. Unlike traditional mesh-dependent models, TripNet supports query-based predictions at arbitrary spatial locations, enabling high-fidelity aerodynamic field estimation without relying on mesh connectivity or downsampled geometry. TripNet achieves state-of-the-art performance on the DrivAerNet and DrivAerNet++ datasets across three major aerodynamic tasks: drag coefficient prediction, surface pressure and wall-shear stress regression, and full 3D volumetric flow-field prediction. TripNet attains an R2 of 0.972 for drag prediction and a relative L2 error of 10.39% for velocity magnitude. TripNet reduces inference time by several orders of magnitude–predicting drag in 0.01 s and full 3D flow fields in 2 s on a single graphics processing unit–while improving accuracy and reducing memory cost compared to strong baselines such as FigConvNet, Transolver, and MeshGraphNet. Beyond in-distribution prediction, we demonstrate that triplane features encode robust, transferable geometric–aerodynamic priors. TripNet exhibits strong cross-shape generalization, outperforming prior works when trained on two car-body classes and tested on an unseen third shape. Furthermore, TripNet achieves data-efficient cross-dataset transfer: when pretrained on DrivAerNet++ and finetuned on the high-fidelity DrivAerML dataset, TripNet surpasses models trained solely on DrivAerML across all metrics, even with as few as 10–50 gradient steps. Collectively, these results highlight triplanes as a scalable, generalizable representation for partial differential equation surrogate modeling and establish TripNet as an accurate, efficient alternative to traditional CFD solvers and existing geometric deep learning models.
SN - 1070-6631
SN - 1089-7666
SN - 1527-2435
ER -