Explainable machine learning based prediction of 28-day compressive strength of recycled aggregate self-compacting concrete using particle packing indicators.

Abstract

This study aims to develop an explainable machine learning framework for predicting the 28-day compressive strength of recycled aggregate self-compacting concrete (RA-SCC) using particle packing indicators. A dataset comprising 365 mixtures from 102 studies was analyzed using six regression models, with SHapley Additive exPlanations (SHAP) applied for interpretability. The Extra Trees model achieved the highest accuracy (R2 = 0.9724, Root Mean Square Error (RMSE) = 1.19 MPa). The dataset is limited to 28-day strength and excludes admixture effects. The study introduces a combined approach integrating particle packing concepts, multicollinearity assessment, and explainable AI. The results demonstrate that machine learning can reliably predict strength while implicitly capturing packing-related mechanisms.