Abstract: To tackle the issues of existing machine learning models, which rely on low-dimensional feature inputs for predicting film-cooling efficiency distributions and lack general applicability, a data-driven framework based on PointNet and conditional generative adversarial network (CGAN) was proposed. In this framework, the irregular mesh nodes within the computational domain were transformed into point cloud data. PointNet was then employed to extract both global and local features from this point cloud. These extracted features serve as conditional information to guide the CGAN generator in reconstructing two-dimensional flow field distributions that closely align with those obtained through computational fluid dynamics (CFD) simulations. Additionally, the adversarial training was conducted by discriminator to improve the detail quality of the generated images. Experiments were carried out through using a sawtooth-shaped channel as the research subject. Results demonstrate that the absolute error between the training and test sets does not surpass 0.05, while the relative error remains below 6%. Moreover, the computational time required is less than that of traditional CFD method. When provided with sufficient training samples, the developed model can be applied to all types of film-cooling structures and effectively address the challenge of accurately reconstructing film-cooling efficiency distributions from sparse point cloud data.