Abstract: Accurate knowledge of the surface pressure distribution is essential for the aerodynamic design and optimization of high-speed trains. However, due to the physical constraints imposed by the train structure, only sparse pressure measurements can be obtained in practical tests. To address this limitation, this study proposes a gradient-guided framework for sensor layout optimization and sparse pressure-field reconstruction. The method integrates a Laplacian regularization based on point-cloud neighborhoods within a deep learning model to enhance spatial smoothness and physical consistency of the reconstructed pressure field. A two-stage training strategy is employed to achieve collaborative optimization of data-driven learning and physical constraints. In the first stage, the network is pretrained to learn the mapping from sparse sensor measurements to dense pressure fields. In the second stage, a differentiable node-selection module is combined with gradient-based constraints to jointly optimize the sensor configuration and reconstruction performance. Experiments conducted on the CRH380A high-speed train dataset demonstrate that the proposed approach achieves RMSE and L2RE values of 1.18×10^–4 and 0.73×10^–3, representing improvements of approximately 4.8% and 3.9% over the PROSNet baseline. The results indicate that the method maintains high reconstruction accuracy and physical fidelity under complex operating conditions, effectively bridging the gap between sparse aerodynamic testing and dense pressure-field acquisition, and providing a reliable tool for aerodynamic analysis and sensor deployment planning.