Microstructural damage dependent machine learning model to predict stiffness reduction in damaged GFRP composites

Abstract

The underlying cause of stiffness degradation in composites subjected to fatigue is irreversible microstructural damage, which exhibits spatial anisotropy depending on the loading conditions and structural characteristics. Each mode of microdamage represents different physical quantities, making it a challenging task to predict stiffness. In this work, machine learning (ML) models, including multiple linear regression (MLR), support vector regression (SVR) and random forest (RF), were employed to predict stiffness based on stereologically quantified damage data. High resolution Scanning Electron Microscopy imaging of edge-sectional planes was conducted during fatigue tests to quantify damage in interrupted composite materials. Experimental findings identified two distinct types of microstructural damage: (i) perpendicular cracks in woven roving mat (WRM) and chopped strand mat (CSM) under tension-tension (T-T) cyclic loading, and (ii) parallel cracks in CSM under both tension-tension and compression-compression cyclic loading. Incorporating these damage features, ML models demonstrated strong predictability of stiffness values for both CSM (T-T) and WRM (T-T) composites, with the SVR model showing particularly good agreement with experimental results for the CSM (T-T) composite. By leveraging experimental microscopy and stereology data, the ML models successfully established a non-linear relationship between microstructural damage and stiffness, providing a robust framework for understanding degradation mechanisms.