Estimation of the permeability coefficient of recycled concrete aggregates using machine learning algorithms: AdaBoost and artificial neural networks

The development of the construction sector in recent decades and the systematic reduction of natural aggregate (NA) resources have obliged the search for alternatives and substitutes for these essential construction materials. Using anthropogenic aggregates as replacements for natural aggregates provides an opportunity to reduce landfills by recycling waste materials from construction and demolition (CDW), for civil engineering applications. It may also involve risks that undoubtedly include estimating geotechnical parameters based on solutions developed to date for natural aggregates. A significant factor affecting these estimates is that anthropogenic aggregates could have different properties and chemical compositions. This study analyzed the possibility of estimating the coefficient of permeability for recycled concrete aggregate (RCA) using machine learning algorithms - AdaBoost and artificial neural networks. The AdaBoost model demonstrated higher accuracy (R2 was 0.959 for training data and 0.886 for test data), relative to neural networks (R2 was 0.682 and 0.640). Shapley Additive Explanations (SHAP) analysis, which increased the interpretability of the results, showed that AdaBoost better reflected the influence of hydraulic properties such as gradient or bulk density, while the neural networks were more sensitive to aggregate grain size. The results indicate the effectiveness of machine learning algorithms in evaluating the coefficient of permeability for recycled materials and represent a promising perspective for supporting the design of sustainable construction solutions.