High-speed and high-fidelity prediction of residual stress field distribution in micro-forging using a physical-translated cGAN

Abstract

Micro-forging (MF) is a surface treatment that induces compressive residual stress (RS) near the surface to improve fatigue performance. However, achieving rapid prediction of RS fields remains a challenging task. In this work, a physical-translated condition generative adversarial network (PT-cGAN) was developed to predict RS fields of the MF process. The PT module translated the non-structured inputs of parameters into nominal RS fields with uniform size, which is suitable for cGAN module. Then the cGAN was trained by the nominal RS fields and finite element (FE) results which used as ground truth. The prediction time of PT-cGAN model has decreased from several hours (FE methods) to a few minutes, with an RS field accuracy (SSIM) of 0.96 and an RS curve accuracy ($R^{²}$) of 0.99. Furthermore, it is attractive to be used for real-time monitoring and parameter optimization.