Abstract: Efficient and accurate prediction of surface flow fields is critical for the multidisciplinary design optimization of hypersonic vehicles. However, conventional CFD methods remain computationally prohibitive for iterative design and lack generalization across varying geometries and flight conditions. To address these limitations, this paper proposes a condition-aware hierarchical point cloud network built upon PointNet++. Specifically, the network integrates a multi-scale grouping set abstraction module to capture rich topological features of vehicle surfaces, along with a geometry-enhanced inverted residual module that incorporates position encoding and channel attention to resolve sharp local flow gradients. Furthermore, to handle the coupling between geometry and operating conditions, a feature-wise linear modulation mechanism is introduced to inject flight parameters as global modulators into the geometric feature stream. The proposed method is validated on a comprehensive hypersonic dataset covering multiple configurations and various operating conditions. As a result, the model achieves average normalized root mean square errors of 0.0124 for surface pressure and 0.0239 for heat flux, with R² exceeding 0.998, and demonstrates robust zero-shot generalization on unseen configurations. Meanwhile, with an average inference time of 115 ms per sample, the framework delivers four orders of magnitude acceleration over CFD, enabling millisecond-level, high-fidelity aerodynamic predictions for next-generation vehicle design and digital twin applications.