Ensemble-based machine learning model to predict the crown of twin-roll thin strip and inverse optimization

Twin-roll strip casting is a near-net-shape manufacturing technology with significant energy-saving and efficiency advantages. However, precise control of strip crown, a key indicator of strip shape, remains a major technical challenge. Conventional mathematical models cannot achieve full-process multi-physics coupling, and few studies have applied data-driven methods to crown prediction. Therefore, this study proposes an ensemble-based machine learning framework for strip crown prediction and inverse optimization of process parameters. A high-quality dataset was constructed from pilot-scale experimental data, and features were refined using metallurgical knowledge, Pearson correlation coefficient analysis, and mutual information comparison. Seven base machine learning models were optimized by Bayesian optimization, and ensemble models including Weighted Averaging, Blending and Stacking were developed to further improve prediction performance. The results show that the ensemble models outperform the base models. The Stacking model achieved the best predictive performance, whereas the Blending model remained competitive with a 16-fold reduction in computational cost. Pairwise statistical comparisons further confirmed that the superiority of Stacking over the other ensemble models remained significant after Holm correction. Considering the trade-off between predictive accuracy and computational efficiency, the Blending model was adopted for interpretability analysis and inverse optimization. SHapley Additive exPlanations analysis quantified the contribution of each process parameter to crown prediction, identifying high-impact features such as t, D_r, W_r and Ra, and low-impact features including Si, Mn, P, Cu, N and Br. A target-driven inverse optimization framework is proposed that uses the target crown and process constraints as inputs and outputs process parameter combinations through the differential evolution algorithm. The inverse optimization framework achieved a relative prediction error of less than 5% in 15 groups of independent experiments, with all optimized parameters falling within the pilot-scale operating range. Offline verification on a pilot-scale experimental line demonstrated stable prediction and inverse optimization capability within the tested operating range. Future work will focus on constructing multi-source industrial datasets and developing online model updating and control integration strategies for adaptive closed-loop control.