Prediction of triaxial behavior of recycled aggregate concrete using multivariable regression and artificial neural network techniques

A cost-effective design of concrete elements in compliance with sustainability principles requires to accurately predict the behavior of concrete under various loading conditions. Yet to date available approaches for the prediction and evaluation of recycled aggregate concretes (RAC) under triaxial load are very scarce, which may hinder the transition of such an eco-friendly concrete from lab-based technique to widespread applications. As a first attempt to iron out this issue, this study develops some reliable and accurate analytical tools on the basis of an experimental database containing the results of 193 cylindrical peak stress under axisymmetric triaxial load, 55 cubic peak stress under true triaxial load, 76 peak strains and 61 elastic moduli of RAC retrieved from an extensive review of the literature. First, a new empirical design-oriented model is presented refining the generic form of conventional confinement models for natural aggregate concrete while considering the impacts of RCAs. Then Grey Correlation Analysis (GCA) is conducted to look into the sensitivity of the key parameters that affect the triaxial behavior of RAC. The results of the GCA indicate that the triaxial behavior of RAC mainly depends on the effective water-to-cement binder ratio, aggregate-to-cement ratio, lateral stress conditions, exposure temperature, as well as the RCA replacement ratio. Having the input parameters identified by the GCA, three robust data-mining mathematic tools, namely Multiple Nonlinear Regression (MNR), Artificial Neural Network (ANN), and hybrid Genetic Algorithm Artificial Neural Network (GA-ANN) are employed to simulate the mechanical performances of RAC under triaxial load. The results demonstrate that the developed MNR equations and the neural network models satisfactorily predict the behavior of RAC under triaxial load; among those models (including the developed design-oriented one), the ANN optimized with GA performs the best. These arguably point to the possibility of the application of those models in the design and evaluation of structural members manufactured using RAC, especially for them under complex loading scenarios.