Abstract
Incorporating waste tyre rubber into cement composites could pose a sustainable strategy for reducing tyre waste, whilst simultaneously preserving natural resources. Accurate prediction of the compressive strength is essential for the efficient development and application of the new material. This study presents the experimental research and a new proposed machine learning modelling framework aimed at predicting the compressive strength of conventional and rubber-modified cement composites containing recycled rubber derived from end-of-life vehicle tyres. Experimental datasets from compressive strength tests on 48 concrete and 40 mortar specimens were used to develop predictive models, with varying rubber content. The empirical results show that increasing the percentage of waste rubber in the mixture designs negatively affects compressive strength, reducing it up to 96.06% for mortars, up to 64.4% for concrete made with ordinary Portland cement, and up to 60.7% for concrete made with sulfate-resistant cement. Three soft computing techniques were implemented using the DTREG predictive modelling software: the General Regression Neural Network (GRNN), Radial Basis Function Neural Network (RBFNN), and Support Vector Machine (SVM). The conventional linear regression (LR) model was also implemented and compared with these three machine learning models. They were trained using the experimentally measured results for the various mixtures. The model performance was evaluated using the standard estimators: the coefficient of determination (R2), correlation coefficient (r), and mean absolute percentage error (MAPE). The GRNN demonstrated the best predictive abilities for predicting the strength of mortars, with MAPE = 4.6% and R2 = 99.27% on the original dataset; on the normalized dataset, MAPE was 3.9%, and R2 was 99.65. For predicting the strength of concrete, GRNN also outperformed the other models on the original dataset with MAPE = 3.8% and R2 = 98.4%. The rubber content was identified as the most influential parameter affecting the compressive strength, while density, admixture conditions, the water-cement ratio, and curing age also contributed to the model’s performance. The LR model presented the lowest predictive accuracy among all evaluated algorithms (models) because of the significant complex nonlinear interactions between predictors and target, which can not be adequately modelled with a linear framework. Within the investigated experimental domain, the developed models showed high predictive accuracy and can be used for preliminary strength estimation and interpolation for mixtures with similar characteristics.