Hot deformation behavior of high-strength non-oriented silicon steel using machine learning-modified constitutive model

Abstract

To overcome the disadvantage of the Arrhenius constitutive (AC) equation in predicting complex nonlinear flow behaviors, this work employed an artificial neural network (ANN) to introduce a proportional coefficient reflecting changes in softening mechanisms under various hot deformation conditions, thereby enhancing the applicability of the AC model. The result demonstrates that, compared to the AC model, the ANN-AC model effectively compensates for errors resulting from different softening mechanisms during hot deformation, with RMSE and MAPE decreasing by 62.47% and 63.75%, respectively. Additionally, the microstructure aligns with the evolution of the power dissipation factor (η) predicted by the ANN-AC model, indicating the map's accuracy. Microstructure analysis shows that discontinuous dynamic recrystallization (DDRX) grains nucleate adjacent to the original grain boundaries at low temperatures and high strain rates, while continuous dynamic recrystallization (CDRX) grains primarily form within original grains at high temperatures and low strain rates. The results reveal the effect of grain orientation on DRX, with //CD (compression direction) oriented grains capable of activating multiple slip systems, indicating a propensity for the CDRX mechanism, whereas //CD oriented grains, with fewer slip systems, promote DDRX through coordinated deformation via grain boundary motion.

Keywords

• Hot deformation
• Constitutive model
• Machine learning
• Dynamic recrystallization mechanism