Bayesian-Optimized Ensemble Models for Predicting Steel Mechanical Properties

Abstract Steel mechanical properties such as yield strength (YS), tensile strength (TS), and elongation (EL) are critical indicators of product quality and safety in downstream applications. In industrial practice, these properties are commonly obtained through tensile testing, which is destructive, time-consuming, and costly, and thus provides limited support for rapid quality feedback in continuous production. This work explores a data-driven alternative that leverages historical production records to predict steel mechanical properties. Motivated by the strongly coupled chemical and physical evolution of steel during heating, rolling, cooling, and heat treatment, we propose an ensemble-learning framework with Bayesian optimization for robust multi-target prediction under real production conditions. Specifically, we construct predictive datasets covering all steel grades produced on the same production line and design a stacking ensemble that uses Random Forest (RF) and CatBoost (CATB) as base learners and Linear Regression (LR) as a meta-learner, with Bayesian optimization for automated hyperparameter tuning. Extensive experiments on independent test data and online production data demonstrate strong predictive performance and practical applicability. Under industrial tolerance bands, the proposed method achieves 92.62% hit rate for YS (±30 MPa), 97.41% for TS (±30 MPa), and 94.53% for EL (±4%), outperforming conventional single-model baselines. These results indicate that the proposed approach can deliver accurate, stable, and generalizable predictions across steel grades, reducing reliance on frequent destructive testing and enabling timely process monitoring and quality control in steel manufacturing.