TOWARDS SUSTAINABILITY: LEVERAGING MACHINE LEARNING FOR GREEN MORTAR FLOW PREDICTIONS WITH WASTE AGGREGATES

Abstract

Accurate prediction of mortar flow is essential for achieving consistent workability, particularly when recycled fine aggregates (RFA) are incorporated due to their higher absorption and angularity. Traditional linear models often fail to capture the non-linear interactions between mixture constituents, motivating the application of machine-learning (ML) techniques. This study applies five ML algorithms Decision Tree, Random Forest, Gradient Boosting, Linear Regression, and AdaBoost to predict the flow of mortars containing varying proportions of natural sand, RFA, and water. In the data set was comprised of 26 samples and mortar flow range was 100 to 270 mm. The prediction was performed only for mortar flow. Exploratory analysis revealed strong positive correlation between water content and flow, while RFA exhibited an expected negative influence. Model performance was assessed using parity plots, error trends, Regression Error Characteristic (REC) curves, and Taylor diagrams. Gradient Boosting consistently delivered the highest predictive accuracy (R-squared = 0.988), showing the closest agreement with experimental flow values, lowest residuals, and largest REC area under the curve. Random Forest and AdaBoost achieved moderate performance, whereas Linear Regression and Decision Tree showed limited generalization capability. The mechanism: boosting sequentially corrects errors of weak learners, capturing non-linear interactions between water, natural sand, and RFA, including absorption and angularity effects. Ensemble averaging reduces variance and handles heteroscedasticity, while tree-based splits adapt to variable interactions and thresholds. Overall, the findings demonstrate that ensemble boosting methods provide a robust and reliable tool for predicting mortar flow, offering significant potential to support mix-design optimization when recycled fine aggregates are used.