Research map / Graph learning / GNN architectures

AeroGTO: An Efficient Graph-Transformer Operator for Learning Large-Scale Aerodynamics of 3D Vehicle Geometries

AAAI 2025 · Graph learning › GNN architectures · approach: Equivariant networks for molecules · 6 citations (OpenAlex)

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 AeroGTO, a graph-transformer neural operator combining a frequency-enhanced kNN-augmented GNN with linear-complexity projection-inspired attention to predict 3D car surface pressure and drag.

Problem

Predicting high-precision 3D vehicle surface pressure and drag coefficient on large, irregular meshes is expensive with CFD, and existing GNN and transformer neural operators struggle with scale, oversmoothing, and cost.

Approach

Combines message passing with sinusoidal positional encoding and kNN edge augmentation for local features, plus a Transformer whose attention projects mesh points onto a small set of learned subspace queries, giving linear complexity in mesh points.

Key results

• AeroGTO achieves mean L2 error 0.0757 on Ahmed-Body and 0.1524 on DrivAerNet, versus 0.0831 and 0.1618 for GINO, a roughly 7.36% average improvement.
• On Ahmed-Body drag prediction, AeroGTO reaches MRE 0.0475 and R2 0.9250, a 10.71% relative error reduction over GINO (0.0532).
• AeroGTO uses 2.442M parameters and 175.298 GFLOPs versus 230.690M parameters and 30839 GFLOPs for GINO.

Measured effects reported in the paper

Surprising findings

• Node-focused sampling reduces FLOPs from 175.3 to 47.7 GFLOPs on Ahmed-Body but raises L2 error from 0.0757 to 0.0896.
• Removing message passing entirely (no MP) gives the largest accuracy drop among ablations (L2 0.1080), while removing kNN augmentation costs less than removing attention.

Limitations

• Drag estimation is pressure drag only, computed from predicted surface pressure
• Only 550 DrivAerNet cars were used due to computational cost, rather than the full dataset
• Ahmed-Body drag results are on a single 51-shape test set

Experimental setup

Datasets: Ahmed-Body DrivAerNet

Benchmarks: Ahmed-Body DrivAerNet

Metrics: mean L2 relative error mean relative error of drag coefficient R2 parameters FLOPs

Baselines: MeshGraphNet GNOT IPOT Transolver GINO

Builds on: MeshGraphNet GINO Transolver GNOT IPOT

Closest papers by meaning

• DragSolver: A Multi-Scale Transformer for Real-World Automotive Drag Coefficient Estimation (ICML 2025 · similarity 0.89)
• Geometry Aware Operator Transformer as an efficient and accurate neural surrogate for PDEs on arbitrary domains (NeurIPS 2025 · similarity 0.87)
• DrivAerStar: An Industrial-Grade CFD Dataset for Vehicle Aerodynamic Optimization (NeurIPS 2025 · similarity 0.87)
• DrivAerNet++: A Large-Scale Multimodal Car Dataset with Computational Fluid Dynamics Simulations and Deep Learning Benchmarks (NeurIPS 2024 · similarity 0.87)
• Aerodynamic Coefficients Prediction via Cross-Attention Fusion and Physical-Informed Training (AAAI 2025 · similarity 0.86)
• Learning Distributions of Complex Fluid Simulations with Diffusion Graph Networks (ICLR 2025 · similarity 0.85)