Integrating physical metallurgy principles and machine learning for predicting continuous cooling phase transformations in steel

Abstract

This study presents a high-precision predictive model for continuous cooling transformation (CCT) behavior in steel by integrating physical metallurgy (PM) principles with machine learning (ML) methods. To enhance industrial applicability, thermal deformation parameters are incorporated alongside chemical composition and cooling conditions. Supplementary features, including dislocation energy (ΔGD) and austenite grain size (Dγ) calculated from PM models, further strengthen the predictive capability. The dataset comprises 64 sets of CCT curves from multiple steel grades. Random forest (RF) and K-nearest neighbor (K-NN) algorithms are used to predict eight key phase transformation temperatures in steels. Model accuracy was validated using both static and dynamic dilatation experiments on 30CrMnSi steel. Evaluated on the internally random-split test subset, the optimized RF and K-NN models achieve a coefficient of determination (R2) at or above 0.99 for most phase transformation temperatures, with mean absolute error (MAE) mostly within 3.0 °C. Exceptions include Ps for the RF model, Ps and Pf for the K-NN model. SHapley Additive exPlanations analysis identifies cooling rate and carbon content as the most influential factors, with PM-derived parameters also playing significant roles. The model performance was further assessed via static and dynamic dilatation experiments on 30CrMnSi steel. Both models accurately capture the rise in transformation temperature and the leftward shift of the CCT curve resulting from deformation. These PM-driven ML models provides an accurate and interpretable tool for predicting static and dynamic CCT curves, supporting intelligent design and process optimization of high-performance steels.