Rapid prediction of vehicle body aerodynamics using an active learning-enhanced graph neural network

Achieving rapid aerodynamic prediction for vehicles under multiple operating conditions is a key challenge in the field of vehicle aerodynamic design. To address this, this study first developed a Graph Neural Network (GNN) surrogate model utilizing a Graph Attention Network (GAT) encoder for force-pressure dual-head prediction. Aerodynamic predictions were performed across the parameter space of various models and operating conditions, and three active learning strategies—uncertainty sampling, diversity sampling and hybrid sampling to the model to improve data efficiency. The results indicate that the GNN-based alternative model performs well in carrying out aerodynamic prediction tasks within the parameter space, with test R 2 values of 0.990 and 0.978 for the force coefficient and pressure distribution, respectively. Active learning strategies significantly improve data utilization efficiency, with the hybrid strategy demonstrating a marked lead in data efficiency ( η 95 = 0 . 22 ), achieving 95% of the performance obtained from training on the full dataset whilst reducing the data volume by 78%. This study successfully applies active learning to the rapid aerodynamic prediction of vehicle-like body models under multiple operating conditions, providing an efficient solution for the rapid iterative development of vehicle aerodynamics.