CT image-based synthetic mesostructure generation for multiscale fracture analysis of concrete

Abstract

Accurate prediction of multiscale fracture process of concrete relies on modeling of concrete mesostructures. Though high-resolution yet realistic mesostructures can be obtained by CT technique, the size limitation of reconstructed mesostructures is still an outstanding task. In this paper, a novel method of generating synthetic mesostructures of concrete based on generative adversarial networks is proposed. After training with CT images of a limited size, a large amount of mesostructures with arbitrary sizes are generated. Subsequently, a systematic scheme for investigating fracture process of concrete is established by incorporating the proposed method with cohesive zone model. The geometric features of generated mesostructures, including two-point correlation functions and aggregate size distributions, are compared with actual CT images to quantitatively verify performance of the proposed method. Simulated crack patterns and macroscale mechanical response of generated mesostructures demonstrate close agreements with those of actual mesostructures, thus validating that the proposed method is an accurate and effective tool in the study of multiscale fracture analysis of concrete.