A deep learning model to classify non-metallic inclusions in steels

• A robust deep learning model established to classify different types of inclusions in steels. • The prediction accuracy of the classification model is high enough. • This methodology could be applied to further study of different aspects of process metallurgy. Deep learning (DL) is the artificial intelligence (AI)-based approach which enables the extraction of complex imaging patterns from the process data and allows for a more accurate prediction in process metallurgy. Compared with the conventional machine learning (ML), this approach enhances the process control efficiency and improves prediction quality. The current study focused on classification of non-metallic inclusions (i.e. Al 2 O 3 and MnS) in a stainless steel as a case study. Several semantic segmentation models were applied to predict the morphology and elemental distribution of MnS and Al 2 O 3 inclusions using images characterized by scanning electron microscopy (SEM) images. The obtained prediction results were compared with inclusion morphology and composition data obtained from energy dispersive spectroscopy (EDS). Finally, OpenCV was used to count the particle size of MnS and Al 2 O 3 inclusion. The prediction results show that the Kernel-based Network (K-Net) model can achieve the highest accuracy for classifying oxide and sulfide inclusions, with a mean intersection over union (mIoU) of 0.8615 and average accuracy (aAcc) of 0.9889 for three-dimensional (3D) images, and an mIoU of 0.8863 and aAcc of 0.9955 for two-dimensional (2D) images. The OpenCV also demonstrated high accuracy in predicting the particle size of inclusions compared with experimental data. The successful implementation of DL approach in non-metallic inclusion classification in steels paves the way for developing intelligence assisted process metallurgy route.