Prediction of thermo-mechanical properties of rubber-modified recycled aggregate concrete: Fid-Bau Portal

Prediction of thermo-mechanical properties of rubber-modified recycled aggregate concrete

Feng, Wanhui / Wang, Yufei / Sun, Junbo / Tang, Yunchao / Wu, Dongxiao / Jiang, Zhiwei / Wang, Jianqun / Wang, Xiangyu

Highlights • The rubber-modified recycled aggregate concrete (RRAC) can be fabricated for low-carbon sustainability. The incorporation of RA and RPs can reduce the strength loss when concrete is exposed to high temperature. • Enhanced peak strain can be obtained through increasing the content of RA and RPs. The properties of RRAC are sensitive to the exposed temperature. • The hyperparameters of machine learning (ML) models were successfully tuned by beetle antennae search (BAS) algorithm. The established ML models possessed high accuracy and good generalisation performance. • The BPNN model possessed highest R value and lowest RMSE value compared to RF, LR, and MLR models, thus it had best prediction performance on the database in this study.

Abstract The recycled aggregate (RA) and waste rubber particles (RPs) can be combined to prepare rubber-modified recycled aggregate concrete (RRAC) effectively contributing to low-carbon sustainability. However, the mechanical characteristics of RRAC must be investigated before the practical application. To this end, this study focused on the uniaxial compressive strength (UCS) and corresponding peak strain of RRAC with versatile design mixtures (i.e. varying contents of RA and RPs) after exposure to different temperatures ranging from 25 °C (room temperature) to 600 °C. The test results exhibited the negative relationship between UCS and RA replacement ratio, RPs content, and temperature. However, RPs positively affected both the loss of UCS and peak strain when RRAC was exposed to high temperatures. Besides, four machine learning (ML) models were developed based on a relatively comprehensive dataset including 120 groups of experimental results. The beetle antennae search (BAS) algorithm was applied to tune the hyperparameter of ML models. The high correlation coefficients (0.9721 for UCS and 0.9441 for peak strain) were determined in modelling using back propagation neural network (BPNN), presenting its accuracy and reliability. Furthermore, BPNN possessed optimal prediction performance since the lower root mean square error (RMSE) and higher correlation coefficient were obtained compared to the other three ML models (random forest, logistic regression, and multiple linear regression).