New machine learning prediction models for compressive strength of concrete modified with glass cullet

Purpose Recycled waste glasses have been widely used in Portland cement and concrete as aggregate or supplementary cementitious material. Compressive strength is one of the most important properties of concrete containing waste glasses, providing information about the loading capacity, pozzolanic reaction and porosity of the mixture. This study aims to propose highly nonlinear models to predict the compressive strength of concrete containing finely ground glass particles. Design/methodology/approach A robust machine leaning method called genetic programming is used the build the compressive strength prediction models. The models are developed using a number of test results on 50-mm mortar cubes containing glass powder according to ASTM C109. Parametric and sensitivity analyses are conducted to evaluate the effect of the predictor variables on the compressive strength. Furthermore, a comparative study is performed to benchmark the proposed models against classical regression models. Findings The derived design equations accurately characterize the compressive strength of concrete with ground glass fillers and remarkably outperform the regression models. A key feature of the proposed models as compared to the previous studies is that they include the simultaneous effect of various parameters such as glass compositions, size distributions, curing age and isothermal temperatures. Parametric and sensitivity analyses indicate that compressive strength is very sensitive to the curing age, curing temperature and particle surface area. Originality/value This study presents accurate machine learning models for the prediction of one of the most important mechanical properties of cementitious mixtures modified by waste glass, i.e. compressive strength. In addition, it provides an insight into the effect of several parameters influencing the compressive strength. From a computing perspective, a robust machine learning technique that overcomes the shortcomings of existing soft computing methods is introduced.