Crystallographic texture prediction of torsioned aluminum wire using hybrid of machine learning and multi-scale crystal plasticity

Abstract

Predicting microstructure after macroscopic deformation is critical for designing materials with specific characteristics and optimized performance. To address the elastic-plastic deformation behavior of materials, computational models based on crystal plasticity (CP) have been extensively developed. However, the high computational cost of CP multi-scale modeling necessitates integrating the CP framework with Machine Learning (ML) techniques. This study combines a multi-scale CP finite element method (CPFEM) framework with ML approaches to predict texture evolution in commercially pure aluminum wire subjected to torsion testing. Initially, the microstructural details of the as-received wire (initial texture), used as input for CPFEM simulations, were determined through electron backscattered diffraction (EBSD) analysis. Subsequently, the multi-scale finite element analysis was performed using commercial Abaqus software and the Düsseldorf Advanced Materials Simulation Kit (DAMASK) code to characterize the microstructural evolution of the samples. Then, two ML techniques were employed to predict texture evolution based on experimental data. The ML models were validated against experimental results obtained from the EBSD analysis, demonstrating their accuracy in forecasting microstructural changes. The comparative analysis of Artificial Neural Network (ANN) and RANSACRegressor techniques revealed that the ANN approach lacks the precision necessary for accurately predicting the texture evolution of 512 seeds. In contrast, the RANSACRegressor method demonstrated significantly higher predictive accuracy, correctly estimating 81.7 % (in degree) of crystal orientations within an error margin of less than 10 degrees. Conversely, the ANN approach achieved only 6.9 % (in degree) accuracy within the same error threshold for the π-radian rotation of the aluminum sample. Furthermore, the results highlight that integrating machine learning techniques with multi-scale CPFEM provides a powerful and innovative framework for predicting the microstructural features of materials, offering significant advantages in computational materials science.