Physical metallurgy guided deep learning for yield strength of hot-rolled steel based on the small labeled dataset

The determination of the relationships between rolling parameters and mechanical properties by high-precision deep learning is important for predicting and controlling the yield strength (YS) of hot rolled steels, however, for most steels, their YS-labeled samples are insufficient for the modeling. In this study, a YS-labeled small dataset and an unlabeled big dataset were firstly collected and then reconstructed the rolling parameters to be microstructures by the physical metallurgical principles. Before being used to label the unlabeled dataset with calculated YS, the strengthening mechanism-based compositions-microstructures-property linkage was optimized by combining the labeled dataset and the particle swarm optimization (PSO) algorithm. To precisely predict YS, the deep neural network (DNN) was initially pre-trained by the big dataset labeled by calculated YS, and then fine-trained by the small dataset labeled by measured YS. Based on it, the effects of compositions, microstructures, and rolling parameters on YS were analyzed, and the results proved to be in good agreement with experimental observations, which presents the solution to implement deep learning by using physical metallurgical principles with a small dataset.