Physics-Attention Wind Noise Transformer: A Point Cloud Deep Learning Surrogate for Rapid Automotive Wind Noise Prediction

Abstract

Accurate prediction of automotive aerodynamic wind noise is important for cabin comfort and early-stage styling, yet conventional CFD and wind-tunnel workflows are too expensive for rapid design iteration. This paper proposes a point cloud surrogate that combines farthest-point sampling with a Transolver-derived, physics-inspired slice-attention mechanism. Here, physics-inspired denotes a representation-level inductive bias; the model does not impose governing-equation residuals, conservation constraints, or physics-based losses. Exterior meshes are converted into 10,240-point geometric inputs and assembled into a controlled dataset of 867 sedan and SUV variants generated at 120 km/h and zero yaw. On the random test split, the model obtains RMSE values of 2.30 dB(A), 2.56 dB for SPL, and 0.0068 for the dimensionless articulation index (AI), with 0.80 s single-sample inference on an RTX 4090. Repeated-seed and grouped-split analyses indicate a favorable accuracy–latency trade-off while also showing a measurable performance decrease for held-out vehicle families. A single-vehicle wind-tunnel comparison confirms strong frequency-trend correlation but reveals a mean simulation over-prediction of 2.70 dB; therefore, the current surrogate should be interpreted primarily as an emulator of the simulation labels rather than a universally unbiased predictor of measured cabin noise.