In the early stages of automotive design, a large number of candidate geometries must be evaluated and their aerodynamic characteristics compared. Traditionally, wind tunnel experiments and computational fluid dynamics (CFD) simulations have been employed for this purpose. However, these approaches involve substantial computational time and cost, and thus become a bottleneck when a large number of geometries need to be evaluated iteratively. In response to this challenge, considerable research effort has recently been devoted to machine-learning-based models for fast prediction of aerodynamic fields. Nevertheless, many representative existing studies focus on a limited set of geometries derived from parametric modifications of a reference vehicle shape. As a result, it remains unclear to what extent prediction accuracy can be maintained for vehicle geometries that were not considered during training. Adapting models to such unseen geometries may require additional large-scale CFD simulations for data generation and subsequent model retraining or significant model adjustment, which can incur substantial computational cost and undermine the practical advantages of machine-learning-based approaches. Therefore, this study aims to investigate the effectiveness of transfer learning in the context of graph neural network (GNN)-based prediction of surface stress fields on automotive bodies. Specifically, we employ XMeshGraphNet[1], which predicts surface pressure and shear stress fields directly from vehicle surface meshes. The model is pretrained primarily on the DrivAerNet++[2],DrivAerML[3], and the pretrained parameters are then transferred to learning tasks involving commercial vehicle geometries[4]. By comparing transfer learning with training without transfer, we quantitatively evaluate the effectiveness of transfer learning in terms of the training time required to reach a target accuracy, prediction accuracy, and prediction stability.