Convolutional neural network for predicting crack pattern and stress-crack width curve of air-void structure in 3D printed concrete

Extrusion-based 3D concrete printing (3DCP) results in deposited materials with complex microstructures that have high porosity and distinct anisotropy. Due to the material heterogeneity and rapid growth of cracks, fracture analysis in these air-void structures is often complex, resulting in a high computational cost. This study proposes a convolutional neural network (CNN)-based methodology for fracture analysis using air-void structures as input. More specifically, the lattice fracture model is used to build a dataset that comprises input air-void structures as well as output fracture information, including the crack patterns and crack-width curves. To establish the relationship between crack morphology and associated microstructures, a U-net convolutional neural network is first presented. With the obtained crack pattern as input, the principal component analysis (PCA) and CNN are then integrated to predict the stress-crack width curves. The predicted results from the CNN model demonstrate a quantitative agreement with lattice numerical analyses, with 0.85 Intersection over Union for crack patterns prediction and 0.75 R2 for the stress-crack width curves prediction. This indicates that CNN models can be used as an alternative to traditional numerical analysis. The feature maps during the convolutional or deconvolutional process are given to explain why the proposed CNN models perform well on fracture analysis of the air-void system. Moreover, the model generalization is discussed using transfer learning with fine-tuning to show the model potential on microstructures expressing varied pore information. In the end, the microstructures cropped from XCT are created to explore the further application of CNN models on fracture analysis of 3D printed materials.