Indentation-informed convolutional neural network for simultaneous prediction of non-equibiaxial residual stress and plastic flow

This study proposes an integrated finite element (FE) simulation and convolutional neural network (CNN) model designed for the simultaneous prediction of plastic properties and surface in-plane non-equibiaxial residual stress from spherical indentation responses. By eliminating the need for stress-free reference specimens, the proposed framework enables non-destructive prediction. The framework leverages indentation load-depth curves and directional deformation profiles derived from validated FE simulations. Sensitivity analyses identify the indenter radius and penetration ratio as critical factors for improving prediction accuracy and maximizing the sensitivity of indentation responses to variations in residual stress. The influence of non-equibiaxial residual stress states on indentation behavior is further elucidated through a mechanistic investigation, which reveals a close association with cumulative volumetric changes in equivalent plastic strain near the indentation zone. The CNN training performance supports the sensitivity-based determination of optimal indentation settings. The model is shown to achieve a mean absolute error corresponding to below 5 % on average for residual stresses, while the plasticity parameters are also well captured. Experimental assessment on copper specimens with homogeneous residual stress fields verifies the accuracy and adaptability of the developed FE–CNN model. Further validation using additively manufactured stainless steel, exhibiting complex heterogeneous residual stresses, shows strong consistency with neutron diffraction measurements. This FE–CNN framework presents a robust and scalable approach for comprehensive mechanical characterization, offering substantial benefits for assessing structural integrity and reliability across diverse industrial applications without recourse to destructive testing.