Recent advances in machine learning have demonstrated strong potential for predicting aerodynamic quantities with high accuracy, although model performance often depends on the availability of large training datasets. In previous work, a systematic analysis was conducted on the prediction of pressure coefficient distributions around two-dimensional airfoils, investigating the influence of Neural Network architectures and training set size using data generated by a potential flow panel method with Kutta condition. Building on these results, the present study extends the analysis to a significantly more complex three-dimensional aerodynamic problem. A Neural Network framework is developed to predict pressure distributions over generic three-dimensional wing geometries, with particular emphasis on identifying the minimum dataset size required to achieve reliable generalization. The increased dimensionality of the geometric and flow parameter space introduces additional challenges in terms of scalability and data efficiency, which are systematically addressed. Since three-dimensional wing configurations can be interpreted as extensions of two-dimensional airfoil sections augmented by additional geometric parameters such as taper and spanwise variation, the previously developed two-dimensional airfoil dataset is used as a reference starting point. This enables an assessment of whether the information contained in the two-dimensional configurations provides a suitable foundation for defining a minimal and effective training dataset for the three-dimensional problem. The results offer quantitative insights into Neural Network generalization capabilities when transitioning from two to three-dimensional aerodynamic configurations and provide useful reference benchmarks for future developments in machine-learning-based aerodynamic modeling.