Identification of damage properties of glass/epoxy laminates using machine learning models

Abstract

In the present research, the possibility of estimating damage properties of Glass/Epoxy woven laminates using machine learning (ML) models is investigated. The dynamic response of laminates under low-velocity impact (LVI) was characterized and used as training data sets for ML models. The suitability of four numerical models was investigated for LVI modeling, and based on their accuracy and simulation run-time, an appropriate method was selected. The MAT054 material model was selected for LVI simulations for two energy states (15 J and 20 J). After characterizing the results of these simulations a set of 13 dynamic characteristics were extracted from the dynamic response of each simulation. Moreover, for each energy state, four ML regression models, namely MLP, KNN, DT, and SVR, were performed with the numerical input sets and the random laminate damage properties in each simulation as the output sets. For each ML model the input vector was the same however only one damage property was set as an output vector. The trained ML models were able to estimate 4 out of 6 damage properties of Glass/Epoxy woven laminates namely Longitudinal modulus, Longitudinal tensile strength, Longitudinal compressive strength, In-Plane shear strength. Additionally, the output of the trained ML models was compared with damage properties obtained from experimental results and it was found that the trained models were able to estimate the said damage properties within 4% – 6% of the experimental results.