A Flow Field Prediction Method Based on the Transformer for Enhancing Local Sensitivity of Irregular Geometries

engineering significance in aerospace, meteorology, and renewable energy systems. While traditional CFD methods face computational inefficiency due to mesh generation and Navier-Stokes equation solving, current deep learning approaches encounter dual challenges with unstructured meshes: voxelization struggles with complex geometries, and PointNet-based methods exhibit limited local feature extraction. To address these, we propose PointTransformerCFD, integrating attention mechanisms to establish learnable local feature aggregation, enhancing physical feature capture in critical regions (e.g., boundary layer separation). Experiments confirm its high accuracy in airfoil flow field prediction.