Interpretable machine learning modeling of alloy composition-process-property relationships based on industrial big data in hot strip rolling

Abstract

Accurate mapping of the composition, process, and property relationship is essential for online predicting and controlling mechanical properties in hot-rolled alloy steel. However, this remains a challenge due to persistent data silos in hot strip rolling (HSR) and the limited interpretability of “black box” machine learning (ML) models in capturing complex multivariable interactions. This study developed a four-layer industrial digital twin platform to integrate multisource heterogeneous data into a unified dataset, including composition, process parameters and properties. A dataset reconstruction strategy was introduced to address the challenges posed by large-scale, nonlinear, and noise-prone data. Based on the reconstructed inputs, interpretable ML models were established to characterize the underlying composition-process-property relationships accurately. The light gradient boosting machine (LGBM) model, optimized using particle swarm optimization, achieved superior performance with an R² of 0.80 and a mean absolute error of 10.02 MPa on the test set. Shapley additive explanations and partial dependence plot analyses further revealed the combined effects of alloying elements, rolling temperature, and deformation on mechanical behavior. The proposed framework was successfully implemented on a 1422 mm HSR production line, providing real-time guidance for alloy design and reducing reliance on manual sampling.