Predicting mechanical properties of steel based on ensemble learning and Bayesian optimization

With the rapid advancement of digital transformation and industrial upgrading in the steel sector, artificial intelligence (AI) technologies have been increasingly implemented in steel production facilities to predict the mechanical properties of steel products. This technological integration aims to reduce the frequency of quality inspections while providing data-driven insights for product quality optimization. However, current predictive models for steel mechanical properties exhibit notable limitations, including suboptimal accuracy and incomplete coverage of all steel grades across the entire production line. To address these challenges, this study proposes an enhanced predictive framework based on ensemble learning combined with Bayesian optimization. The model architecture employs Random Forest (RF) and CatBoost (CATB) as base learners, Linear Regression (LR) as a meta-learner, following a stacking strategy to improve predictive performance. Comprehensive experimental evaluations demonstrate that the proposed ensemble model significantly outperforms conventional single machine learning approaches. On independent test datasets, the prediction accuracy of the Yield Strength (YS) prediction model is 92.62% (± 30 MPa absolute error), the prediction accuracy of the Tensile Strength (TS) prediction model is 97.41% (± 30 MPa absolute error), and the prediction accuracy of Elongation (EL) prediction model is 94.53% (± 4% absolute error), representing a substantial improvement over existing methods. The results highlight the superior predictive capability of the ensemble learning framework, underscoring its practical significance and potential for widespread adoption in industrial applications. This advancement not only enhances prediction accuracy but also ensures comprehensive coverage of all steel grades, thereby contributing to more efficient quality control in steel manufacturing.