ABSTRACT In this study, a machine learning-based defect detection framework is proposed to estimate three dimensional (3D) defect spatial information in complex-shaped carbon fiber reinforced plastics (CFRPs) using surface stress distributions obtained from infrared stress measurement [1]. While CFRPs are essential for complex structures like prosthetic legs, detecting internal defects in such curvilinear geometries remains challenging for conventional methods. The proposed method utilizes a graph neural network (GNN) capable of learning from 3D mesh data. A major challenge in this approach is the scarcity of experimental data, which prohibits the direct training of the GNN. To address this, the GNN is trained exclusively using a large dataset generated by homogenization finite element analysis (FEA). To enable the FEA-trained GNN to process experimental data accurately, we introduce a statistical domain adaptation framework integrating principal component analysis (PCA), partial least squares (PLS), and gaussian process regression (GPR). Specifically, PCA is first applied to the FEA stress distributions to derive their latent variables. Subsequently, PLS is employed to extract features from the experimental infrared images that are statistically correlated with the FEA physics, filtering out noise. Finally, GPR constructs a non-linear mapping function that connects these PLS-extracted experimental latent variables to the PCA-reduced FEA latent variables. This process effectively aligns the experimental input with the simulation-based feature space. The framework was validated through cross-validation using experimental data from a CFRP prosthetic leg with artificial defects. The results demonstrated that the proposed method successfully predicted the presence and approximate 3D geometry of internal defects from experimental images, effectively overcoming the data scarcity issue. REFERENCES [1] Offermann, S., Beaudoin, J. L., Bissieux, C., and Frick, H., Thermoelastic stress analysis under nonadiabatic conditions, Experimental Mechanics, 37(4), 409–413, 1997.