Machine Learning-Based prediction of flexural strength in recycledaggregate concrete: a K-Nearest Neighbors approach

Abstract

Predicting the flexural strength of recycled aggregate concrete (RAC) is challenging because recycled constituents introduce variability and nonlinear interactions. This study develops a K-nearest neighbors (KNN) model to predict 90-day flexural strength using 366 mix designs collected from peer-reviewed studies published between 2010 and 2024. Eighteen mixture, supplementary cementitious material, and chemical-modulus variables were considered. After preprocessing and hyperparameter optimization, the model achieved R2 = 0.812 on the independent test set. SHAP analysis identified the water-to-binder ratio, natural coarse aggregate content, and maximum aggregate size as the most influential predictors. Partial dependence plots further revealed nonlinear effects and interactions among key variables. Most predictions were within ±10% of experimental values. The results show that KNN can support preliminary RAC mixture screening and provide interpretable guidance for optimizing flexural performance while incorporating recycled aggregates.