Application of deep learning for predicting carbon segregation in large-diameter continuously cast steel billets

Carbon segregation is a persistent defect in continuous casting of large-diameter steel billets, leading to deteriorated mechanical properties and compromised service reliability. Conventional empirical or machine learning models generally estimate segregation indices but cannot resolve local variations of carbon distribution across billet sections. In this work, a microstructure-informed convolutional neural network (CNN) framework is proposed to predict and map carbon segregation in 600 mm round 42CrMo steel billets. A comprehensive dataset comprising microstructural images and corresponding carbon content measurements was established. The customized CNN achieved a testing accuracy of 81.3% with a mean absolute error of 0.012 wt.% and showed good robustness in out-of-sample validation. Compared with transfer learning models (VGG16, VGG19, etc.), the customized architecture exhibited superior generalization on this domain-specific dataset. Contrast-enhanced imaging significantly improved predictive performance, while Grad-CAM visualizations highlighted key microstructural regions correlated with carbon distribution, providing interpretability. This study demonstrates a proof-of-concept methodology to achieve quantitative mapping of segregation patterns in large-diameter 42CrMo billets, offering a complementary tool to traditional metallurgical analysis and providing a workflow that may support future data-driven research on segregation formation mechanisms and process optimization in steel casting when extended to additional steels and casting conditions.