Machine learning prediction of fiber pull-out and bond-slip in fiber-reinforced cementitious composites

Abstract

Single fiber pull-out and fiber-matrix interfacial interaction play an essential role in understanding the mechanical behavior of fiber-reinforced cementitious composites. The present study introduces a computational model for predicting the maximum fiber pull-out force and corresponding bond slip. An extensive literature survey was performed to create a pertinent comprehensive experimental database. A total of 382 experimental data were utilized to develop and train the Artificial Neural Network (ANN) models. The model input parameters included the fiber embedded length, fiber inclination angle, fiber tensile strength, fiber length-to-diameter ratio, loading rate, water-to-cement ratio, concrete compressive strength, and fiber geometry. The model output consisted of the maximum pull-out force and corresponding slip. The results indicate that ANN with two hidden layers and 12 neurons was adequate for predicting the outputs with a mean absolute percentage error (MAPE) of less than 10%. To obtain the importance of the inputs on the outputs (the maximum fiber pull-out force and the corresponding slip), a sensitivity analysis was done based on the Milne formula on the proposed ANN. According to the results, it was found that among the eight inputs, the parameters of the geometric shape of the fibers (straight, hooked-end and spiral fibers) and fiber tensile strength have the highest effect on the outputs, with an impact percentage of 16.1 and 15.1, respectively. The mean square error (MSE) was 0.9 for the maximum pull-out force and 0.14 for slip, respectively. Overall, the proposed executed model attained reasonable predictions and could offer a data driven approach to optimizing fiber-reinforced cementitious composites.