A novel microstructure-informed machine learning framework for mechanical property evaluation of SiCf/Ti composites

Abstract

In this work, a novel microstructure-informed machine learning (MIML) framework is proposed for the longitudinal ultimate tensile strength (UTS) evaluation and the construction of microstructure-property linkages of titanium matrix composites (TMCs). The main components of the MIML framework include (i) high-fidelity simulation and dataset generation using finite element method (FEM), (ii) the quantitative characterization of microstructures obtained by 2-point cross-correlation analysis and principal component analysis, (iii) the extraction of structure-property linkages based on support vector regression model and (iv) the validation of MIML by experimental data. The fiber breakage observed by advanced microstructure characterization is considered as the dominant failure mode during mechanical tests of TMCs, which is embedded in the FEM simulations. The MIML approach reaches capable accuracy in UTS prediction of TMCs using FEM-generated datasets and trained MIML is validated by experimental data. In addition, the permutation importance-based method is utilized to accurately identify the key feature parameters related to UTS. The first two principal components of microstructural reduced-order representation are recognized as highly relevant features for UTS. The well-trained MIML model is a promising tool in the evaluation of the mechanical property, and further provides a technical chain for the establishment of microstructure-property linkages of multi-phase composites.

Keywords

• Microstructure-informed machine learning
• Ultimate tensile strength
• Titanium matrix composites
• Two-point correlation
• Principal component analysis