On developing accurate prediction models for residual tensile strength of GFRP bars under alkaline-concrete environment using a combined ensemble machine learning methods

Abstract

GFRP (Glass-fiber reinforced polymer) bars are recognized as a structural material enabling replace existing steel rebar. However, GFRP bars exhibit a decrease in tensile strength under severe conditions such as strong alkalinity, high salinity, and humid environment. Thus, a predictive model for such GFRP tensile strength deterioration attempts to be developed, but model accuracy still needs improvement. Therefore, this paper proposes a more enhanced ensemble machine learning model to predict the residual tensile strength of GFRP bars accurately. For this end, tensile strength retention (TSR) experiment results of GFRP bars are utilized. Critical parameters for GFRP TSR are diameter, fiber volume fraction, pH, temperature, and exposure time. Regarding the TSR prediction model of GFRP bar, single machine learning models such as multiple linear regression, nonlinear regression, support vector machine, artificial neural network, and Gaussian process regression show 0.482–0.894 for training and 0.412–0.813 for testing, based on the accuracy of coefficient of determination (R2). Individual ensemble learning machine learning models of bagging and stacking show an accuracy of about 0.897 for training and 0.816 for testing. The proposed model shows an accuracy of about 0.912 for training and 0.834 for testing, which improves about 4–22% compared to previous study model performances.