In this study, a machine learning model is developed for prediction of three-dimensional information of defect from two-dimensional stress distribution of carbon fiber reinforced plastic (CFRP). A model of prosthetic leg made by CFRP is chosen. The graph neural network or transformer is employed. Both experimental data and numerical analysis results are used as the training data of stress distribution. Infrared stress analysis is used for obtaining the experimental data. The finite element analysis is performed to obtain the simulation results. Especially, the homogenized finite element analyses are conducted for the plain weave microstructure and unidirectional microstructure.