Study of recycled concrete properties and prediction using machine learning methods

Abstract

This paper is divided into two parts, with the first part examining the influence of recycled brick aggregate (BA) ratios on physical properties of blended coarse aggregate (BNA). These properties include apparent density (SSD), water absorption (WA), and crushing value (CV). The compressive strength of recycled concrete (RAC) prepared with BNA was tested. The tested results show that when SSD decreases from 2700 kg/m3 to 2000 kg/m3, the compressive strength of the corresponding RAC decreases by 40.5 %. Moreover, when the WA increases from 5 % to 10 %, the compressive strength of the concrete decreases by 15.9 %. Furthermore, as the CV rises from 9 % to 37 %, the compressive strength of the concrete decreases by 38.4 %. And the second part explores the application of machine learning models to predict the concrete compressive strength, incorporating parameters such as BA replacement rate, SSD, WA, CV, water-to-cement ratio (W/C), and blended coarse aggregate to cement ratio (BNA/C). To improve model accuracy, we integrated data from our own experimental tests with an additional 128 datasets sourced from literature, resulting in a combined dataset of 133 entries for training and testing the machine learning models. Three machine learning models were proposed and compared: the BP neural network, GA-optimized BP (GA-BP) neural network, and convolutional neural network (CNN) model. Compared to the other two models, the CNN model exhibits the highest prediction accuracy and generalization ability, as evidence by its training accuracy of 0.969 and test accuracy of 0.927. The CNN model proved capable of facilitating intelligent RAC mix design in engineering applications. A case study was presented to illustrate the utilization of the CNN model in predicting the W/C and BNA/C ratios for 11 different BA ratios, ensuring the targeted RAC strength is achieved with specific type of BNA.