Prediction of residual stress in additively manufactured 18Ni300 maraging steel using GAN-based deep learning method

To optimize LPBF-PHT process parameters for minimizing residual stress, a deep learning–based residual stress prediction model was integrated with the non-dominated sorting genetic algorithm (NSGA-II). The model, built upon StyleGAN2-ADA, predicts residual stress distribution images before and after PHT from the process parameters, including laser power, scan speed, annealing temperature, and holding time. A high coefficient of determination ( = 95.96%) was obtained using only 45 non-augmented samples through sensitivity analysis of augmentation strategies suitable for residual stress fields within the ADA method. The model captured the physically trend of residual stress mitigation with increased energy density even under unseen process conditions. Additionally, SHAP analysis revealed that stress reduction was primarily governed by the annealing temperature and was further enhanced when the process conditions resulted in higher energy density, such as higher laser power and lower scan speed. Accordingly, the Pareto optimal solutions obtained through NSGA-II were concentrated in the high energy density region.