Data-driven quantification of orientation dependent damage caused by voids using Machine Learning

Abstract

Voids have a significant impact on the structural safety and performance of polycrystalline metal alloys given their crucial role on the initiation and evolution of damage. Therefore, a fundamental understanding of the relationship between the internal crystalline structure of metal alloys and their corresponding damage behavior and properties is essential for the materials community. Crystal plasticity theories, in conjunction with finite element (CPFEM), are actively used to describe and characterize this behavior given the fact that they directly consider the orientation of the crystallographic plains, slip systems and other microstructural features. Nevertheless, despite its accuracy, CPFEM-based analysis protocols are often ill-suited for establishing a computationally efficient and accurate linkage between the microstructure and the resulting damage performance given their high computational cost and their need to iteratively solve complex, numerically stiff and highly non-linear equations. In this work, we address this challenge by establishing a machine learning (ML)-based linkage between the microstructure and the resulting damage performance. Specifically, we leverage AdaBoosted decision trees to connect crystal orientations, represented with Generalized Spherical Harmonics, to a measure of damage derived from the classical Lemaitre continuum damage model. The developed ML model accurately predicts the Lemaitre stress around a spherical void at a fraction of the computational cost compared to CPFEM simulations.