Prediction of residual stress field of ultrasonic rolling processed 20CrNiMo based on physics-informed neural networks

Abstract

The ultrasonic rolling (UR) process enhances the residual stress (RS) in 20CrNiMo, thereby improving its fatigue performance. A comprehensive understanding of the RS distribution under varying process parameters is essential for the optimal application of UR, but the RS evaluation typically relies on extensive experiments and simulations, which are time-consuming and laborious. This paper presents a physics-informed neural network (PINN) model aimed at precisely predicting the distribution of RS fields in ultrasonic rolled 20CrNiMo, addressing the existing challenges. First, a combined function using Gaussian and exponential components is formulated, based on established laws, to depict the RS distribution along the depth direction. Then, a model framework based on PINN is designed, where the combined function is embedded into the hidden layer of the neural network as physical neurons to guide the training process. The undetermined parameters of the combined function are estimated by another hidden layer of the PINN. Finally, by incorporating RS field characteristics, a physics-based loss function is designed to improve the model's generalization and accuracy in predictions. Embed combined function in the hidden layers and incorporate physical constraints into the loss function, thereby transforming a traditional ANN model into a PINN model. The verification results indicate that the PINN model demonstrates excellent predictive performance, with an R 2 of 0.92 and an RMSE of 110.97 MPa on the test set, an improvement of 0.06 in R 2 and a reduction of 35.25 MPa in RMSE compared to the ANN. Additionally, the PINN outperforms the traditional ANN model in predicting the location of maximum RS. This study's PINN model offers an effective solution for refining UR parameters and boosting the fatigue resistance of 20CrNiMo transmission components.