Abstract: The rapid advancement of intelligent technology has brought opportunities for innovation in the computational modeling of aircraft aerodynamic performance. This paper develops an intelligent prediction method for aerodynamic performance based on point cloud representation and Transformer architecture for three-dimensional aircraft, achieving end-to-end prediction of aerodynamic parameters and accelerating the process of solving aerodynamic problems. Furthermore, to address the contradiction between the large sample requirements of intelligent models and the small sample characteristics of the aerodynamic field, a transfer learning framework for aerodynamic performance prediction tasks is developed based on model parameter reuse methods and domain adversarial learning, effectively reducing the training sample requirements of the model. In terms of dataset construction, this paper builds a generalized test dataset containing over ten thousand samples, including shape generalization, condition generalization, and dual generalization, based on proprietary simulation performance data of three-dimensional aircraft. The results show that the aerodynamic performance prediction model requires only 3.096 seconds for predicting the aerodynamic performance of a single three-dimensional aircraft, achieving a computational efficiency improvement of two orders of magnitude compared to computational fluid dynamics (CFD) methods. The non-standard mean relative error of the prediction results is only 5.6%, and the model demonstrates excellent generalization across shape and condition test datasets. Under the transfer learning framework proposed in this paper, the aerodynamic performance prediction model can maintain the same accuracy level (10%) while reducing the training sample requirement by 70%, effectively addressing the challenge of scarce three-dimensional simulation data in aerodynamics. This approach is expected to provide key technical support for aircraft design optimization and rapid development.