Speaker
Description
Accurate rolling force prediction in flat hot rolling is essential for precise gap control, shape stability, and equipment protection, operating within mill power and force limitations. In practical hot rolling operations, the prediction of rolling forces is challenged by the combined effects of evolving deformation geometry and different metallurgical phenomena such as recrystallization, recovery, strain accumulation, precipitation of particles and thru-thickness temperature evolution, resulting in different material responses from pass to pass.
This work proposes a physics-informed hybrid machine learning framework that combines production process raw data with metallurgical features computed by a microstructural simulation tool (MicroSim®) to predict rolling forces. Feature engineering is structured in two layers: (i) physically derived variables, including contact length, and shape factor, and (ii) pass-level with austenitic grain size distribution, recrystallized fraction, accumulated strain, and metallurgical mean flow stress (MFS). The MFS acts as a physics anchor, providing the model with theoretical baseline resistance of the material and shifting the learning task toward correcting deviations from physical metallurgy.
A gradient boosting algorithm (XGBoost) was benchmarked, across an industrial dataset, in two configurations: a pure data-driven baseline and the proposed hybrid model. The hybrid model consistently outperformed the baseline in accuracy and generalization, with feature importance analysis confirming that microstructural variables ranked among the dominant predictors, validating the physical consistency of the approach.
These findings indicate that combining process data with thermomechanical simulation outputs provides a robust tool for rolling-process prediction, directly supporting improved control and enhanced product quality.