Improved design method for gas carburizing process through data-driven and physical information

Abstract

Data- or physics-driven computational simulation algorithms have gained widespread attention in the field of scientific computing. However, most existing methods rely solely on either data or physical information to solve problems, making them susceptible to the complexities of physical processes or issues such as data loss and distortion. In this paper, we propose a dual-driven simulation method that combines data and physical information to improve the accuracy and stability of carburizing heat treatment simulations, specifically addressing the data loss problem faced by purely data-driven models. By embedding Fick’s second law into the deep learning framework, we created a Physics-Informed Neural Network (PINN) to simulate the transfer and diffusion of carbon elements in the carburizing process. This method breaks the neural network’s dependence on data. Based on this, we developed an efficient gas carburizing process design method and validated its accuracy and efficiency on typical carburizing steel, with a deviation of only 0.008% from the target carbon concentration. In terms of neural network solver design, we optimized and discussed the network’s hyper-parameters, finding that a network design with three hidden layers offers the best accuracy for this type of problem without imposing a heavy computational burden. Compared to classical numerical solvers, this method increases computational speed by several orders of magnitude.