Machine learning and materials informatics approaches for predicting transverse mechanical properties of unidirectional CFRP composites with microvoids

Abstract

The mechanical properties of composites are traditionally measured using numerical and experimental approaches, which impede the innovation of materials due to the cost, time, or effort involved. This study for the first time develops a machine learning-assisted model, which aims at predicting the transverse mechanical properties of unidirectional (UD) carbon fiber reinforced polymer (CFRP) composites with microvoids. To this end, the stochastic microstructures are generated by random sequential expansion (RSE) and hard-core model. And the transverse elastic modulus, transverse tensile strength and transverse compressive strength are computed by micromechanics-based finite element (FE) method. Then, the reduced order representation of microstructures is determined using 2-point spatial correlations and principal component analysis (PCA). Finally, a genetic algorithm (GA) optimized back propagation (BP) neural network is implemented to capture the potential nonlinear relationship between microstructure and transverse mechanical properties. The presented data-driven techniques can reproduce the FE simulation results with an R-value of 0.89 or greater. The excellent agreement between the predicted results and test datasets verifies the successful application of data science methodologies in elucidating the microstructure-property linkages of UD-CFRP composites with microvoids, and thus provides a promising tool for accelerating the smart design and optimization of composites.