Deep learning model for predicting hardness distribution in laser heat treatment of AISI H13 tool steel

Abstract

In this paper, a deep learning model for predicting a hardness distribution in laser heat treatment of AISI H13 tool steel is presented. As an input, this model uses a cross-sectional temperature distribution obtained from a 3-D thermal simulation and transforms it to a hardness distribution. Unlike other physics-based predictive models, where a complete temperature history during the entire heat treatment process is necessary, with this model, hardness prediction was possible from only one temperature distribution that is obtained when the surface temperature reaches a maximum value. A 2 kW multi-mode fiber laser was used to heat-treat steel specimens, and measured hardness distributions were used as ground truths for the model. The model was constructed based on a conditional generative adversarial network (cGAN) architecture with a convolutional neural network (CNN). Although a training data set consisting of only three process conditions was used for the supervised learning, input temperature distribution images were successively translated to the corresponding hardness distribution images. The average accuracy of the predictions was 94.4%, which is roughly 10% better than the prediction accuracy of the authors’ carbon diffusion time model (Oh and Ki, 2017).