Deep-learning model for predicting hardness and phase distributions from two cross-sectional temperature distribution images in laser heat treatment of AH36 steel

Abstract

Laser heat treatment of carbon steel is generally performed to increase the hardness of the specimen. However, when the heating temperature is high, softening of the specimen can occur along with melting. It is important to predict both hardening and softening processes, but such research has been limited so far. In this study, a deep learning model was developed that predicts both hardening and softening during laser heat treatment of AH36 steel using two cross-sectional temperature distributions obtained at 0.5-s intervals as inputs. Two temperature distributions were used as inputs to provide accurate information about cooling rate to the model. The input temperature distribution of the cross-section was calculated by solving the heat conduction equation using finite element method, and the hardness and phase distributions used as ground truth were obtained through a 2 kW multi-mode fiber laser experiment. The model was constructed based on a generative adversarial network with a residual network. It was found that when two temperature distributions were used, both softened and hardened areas were predicted accurately. However, when only one temperature distribution was used, only hardening could be predicted. On the other hand, both models could predict phase distributions, but much more accurate results were obtained when two temperature distributions were used as inputs.