From compressive strength studies to predictive machine learning models: Rubberised concrete containing brick powder

Abstract

Through waste tire rubber recycling and thanks to very simple pozzolanic materials such as brick powder (BP), it has been demonstrated that sustainable construction can be achieved during concrete production. Since concrete is a heterogeneous material with variable and complex behaviour by nature, it is important to incorporate machine learning (ML) models in forecasting its behaviour. Although ML models have been employed for predicting concrete containing BP and/or tire rubber aggregate (TRA), no studies have explored the use of adaptive boosting (AdaBoost), light gradient boosting machine (LightGBM), extreme gradient boosting (XGBoost), gradient boosting regression (GBR), cluster regression, multilayer perceptron (MLP) and Gaussian process (GP) models to forecast the behaviour of rubberised concrete containing BP. In this comprehensive research, the foregoing ML algorithms are employed to forecast the compressive strength of rubberised concrete with BP. The findings illustrate that the GBR model is superior during predictions for the training, validation and testing stages, as evidenced by higher R² values ranging from 0.77 to 0.98. SHarpley Additive exPlanations (SHAP) analysis results reward age as the highest influential variable having an average SHAP value of 3.561, followed by tire rubber aggregate, coarse aggregate and cement. In addition, pronounced model performance differences are observed using the Taylor diagram analysis. The research also establishes a predominantly overfitting behaviour displayed by most folds during k-fold cross-validation. Regularisation of the model is proposed to prevent overfitting by penalising model complexity. The ML algorithms are competent to predict the compressive strength of rubberised concrete with BP well, thereby enabling practitioners and engineers to make versatile decisions regarding concrete mix designs and quality controls.