A point cloud network-embedded multi-fidelity surrogate model for fast aerothermal prediction of hypersonic vehicles with variable configurations

Fast and accurate aerothermal prediction is essential for hypersonic vehicles. In this study, a Point Cloud-Embedded Multi-Fidelity Network (PC-MFNet) is proposed for aerothermal prediction with variable configurations. Within this framework, low-fidelity heat flux data generated by Eckert's reference enthalpy method are embedded into a point cloud neural network, enabling effective multi-fidelity data fusion to enhance prediction accuracy and computational efficiency. To construct the multi-fidelity dataset, four representative hypersonic configurations—the double ellipsoid, blunt cone, blunt biconic, and lifting body—were selected based on computational fluid dynamics (CFD) simulations and engineering calculations. The generalization performance of PC-MFNet was evaluated through multi-dimensional test cases with totally different configurations. Results show that PC-MFNet maintains prediction errors below 4% under various angles of attack and Mach numbers for the trained configurations. For unseen configurations outside the training set, the model achieves average prediction errors for heat flux below 14%, demonstrating strong generalization performance. Moreover, PC-MFNet requires only 0.2% of the prediction time compared to CFD simulations while maintaining near-CFD-level accuracy.