Prediction of tensile properties in fiber-reinforced composites by machine learning models
TL;DR
The results validate the feasibility of applying ML, particularly instance-based and ensemble learning models, to predict the tensile behavior of fiber-reinforced composites using a limited experimental dataset, thereby reducing reliance on destructive testing and accelerating the design of high-performance hybrid composites for aerospace and automotive applications.
Abstract
The consistent design of fiber-reinforced composites (FRCs) remains challenging due to their heterogeneous structures and the complex relationships among processing parameters and material properties. Conventional tensile characterization is costly and destructive, while existing predictive models are often limited to specific tensile parameters or single-fiber composite systems. This study addresses these limitations by developing machine learning (ML) models to simultaneously predict multiple tensile properties of both pure and hybrid composites using experimental data. A total of 54 laminates were fabricated and subjected to tensile testing, including pure and hybrid compositions of carbon, Kevlar, and glass fibers in cross-ply and quasi-isotropic lay-up with 4-, 8-, and 12-ply configurations. Several ML algorithms were implemented, including baseline algorithms [Linear Regression (LR), K-Nearest Neighbor (KNN), Decision Tree (DT), and Support Vector Machine (SVM)], ensemble models [Random Forest (RF), Gradient Boosting (GB), Stochastic Gradient Boosting (SGB), Extreme Gradient Boosting (XGBoost), and Adaptive Boosting (AdaBoost)], and Artificial Neural Network (ANN). The experimental results showed that the Kevlar-glass bi-hybrid cross-ply 4-ply laminate (373.46 MPa) and the carbon-glass-Kevlar tri-hybrid cross-ply 8-ply laminate (389.79 MPa) exhibited higher tensile strength than the pure Kevlar laminate (326.40 MPa). Among the evaluated models, K-Nearest Neighbor (KNN) achieved the best predictive performance (R2 = 0.87, MSE = 318.84, MAE = 7.73), followed by Stochastic Gradient Boosting (SGB) (R2 = 0.78), while the ANN demonstrated moderate performance (R2 = 0.41). The results validate the feasibility of applying ML, particularly instance-based and ensemble learning models, to predict the tensile behavior of fiber-reinforced composites using a limited experimental dataset, thereby reducing reliance on destructive testing and accelerating the design of high-performance hybrid composites for aerospace and automotive applications.