XGBoost and CNN-based prediction of external forces and stress-strain during solidification reduction in casting

Abstract

The continuous casting reduction process can effectively improve defects such as central segregation and porosity in the billets, making it a key process for enhancing billets quality. This study aims to investigate the required external force during the solidification reduction process and the equivalent stress-strain distribution at various positions inside the billets, and to construct predictive models for equivalent stress-strain and external force based on the reduction amount. Through conducting laboratory ingot solidification process reduction tests and finite element numerical simulations, the corresponding parameters including external force, reduction amount, equivalent stress, and equivalent strain were obtained. Data augmentation and random noise processing were applied to the parameter features, and predictive models for external force using the XGBoost algorithm and for equivalent stress-strain using a convolutional neural networks were constructed. The results indicate that as the reduction amount increases, the required external force increases, and the equivalent stress-strain inside the billet is much higher than at the surface. Furthermore, in the external force prediction model, the MAPE value of the XGBoost algorithm is 8.15 ​%, while in the equivalent stress-strain model, the R² value of the convolutional neural networks training set is 0.92. These results demonstrate that both the external force prediction model and the equivalent stress-strain prediction model exhibit high accuracy, and robustness in predictive tasks.