Explainable deep learning-aided design of alloyed carbon steel with improved hot ductility for manufacturing scalability

• Deep learning model predicts hot ductility from composition and casting conditions. • Model captures six ductility curve shapes with high accuracy (MAE = 5.11%) • Temperature is the strongest factor; niobium and carbon are influential • Explainability shows niobium lowers ductility; titanium–nitrogen balance matters. • Case study reduces surface cracks and supports industrial casting decision. This study proposes an explainable deep neural network model to predict the reduction of area (RA) behavior of diverse steel compositions under various casting conditions, and demonstrates its application to the design of new alloyed carbon steels that have increased castability. The model was trained on 4,528 RA data collected from in-house experiments and the literature. It incorporated 16 steel components and 6 thermal history variables. It achieved high prediction accuracy (RMSE = 7.16%, MAE = 5.11%, R2 = 0.91) and successfully described six distinct RA curve types. The use of Shapley Additive Explanations (SHAP) to evaluate the contribution of each input variable enabled metallurgical interpretation of individual RA predictions. The SHAP analysis revealed complex interactions among key elements (e.g., Nb, Ti, N), and these elements have a negative impact on RA value when their contents increase. A case study highlighted the model’s utility in the design of a new alloy for automotive applications. This approach can avoid the disadvantages of conventional trial-and-error methods and provides a scalable, cost-effective method for predicting hot ductility and designing alloys for industrial continuous casting.