ANN-based prediction of bainite start temperature in Fe-C-Mn-Si-Cr-Ni-Mo steels: comparison with empirical models and metallurgical insights

Abstract

This work presents an interpretable artificial neural network (ANN) model for precise prediction of the bainite start temperature in Fe-C-Mn-Si-Cr-Ni-Mo steels, trained on 46 experimentally validated compositions. The model captures the complex, nonlinear influence of multiple alloying elements and surpasses nine well-established empirical equations, achieving a Pearson's R of 0.987 and a mean absolute error of 8.25 degrees C. Through comprehensive sensitivity analysis, the roles of individual elements and their interactions are quantified, identifying manganese and carbon as the most impactful. Ternary contour diagrams offer intuitive visualization of austenite and ferrite stabilizer effects, enhancing metallurgical interpretation. The model's reliability is confirmed through independent validation, with prediction errors under 4%. To facilitate practical application, a user-friendly graphical interface enables real-time property prediction, virtual alloy design, and trend exploration. This ANN-based approach integrates accuracy, insight, and accessibility-offering a powerful tool to accelerate alloy development while minimizing experimental cost and effort. The model is limited to the reported composition ranges and thermal conditions, and future work will expand the dataset and employ DoA and XAI methods to improve applicability and interpretability.