Physics-informed machine learning with residual learning for width spread prediction and diagnostics in hot rolling steel industry

Width spread is a critical quality indicator in the hot strip rolling (HSR) manufacturing process. To improve prediction accuracy, a physics-informed machine learning framework with residual learning (PI-MLRL) is proposed, in which a mechanism model, a light gradient boosting machine (LightGBM)-based residual learning module, and a physics-constrained distillation mechanism are integrated. By combining physical consistency with nonlinear fitting capability, an accurate mapping between process variables and width spread is achieved. Experimental results show that the proposed framework outperforms the mechanism model and seven representative data-driven models in terms of mean absolute error, root-mean-square error, and coefficient of determination. Moreover, Shapley additive explanations (SHAP) method is employed for interpretable diagnostics of PI-MLRL predictions, clarifying the effects of key variables on width spread under different operating conditions. Finally, the proposed framework was deployed on a 2160-mm HSR production line, and application results showed that the width spread prediction error was maintained within ± 3 mm, thereby confirming its engineering applicability.