Machine Learning-Based Prediction of Compressive Strength of Recycled Aggregate Concrete

Abstract

Recycled aggregate concrete (RAC) provides a promising route for reducing the consumption of natural aggregates and diverting construction and demolition waste from disposal. However, the prediction of RAC compressive strength remains challenging because it is governed by the combined effects of mixture proportions, recycled aggregate characteristics, parent concrete quality and aggregate morphology. The present study develops a comprehensive author-curated database of 637 RAC observations and investigates the applicability of two ensemble machine-learning algorithms, Random Forest (RF) and Extreme Gradient Boosting (XGBoost), for predicting compressive strength. Twelve input parameters were considered: effective water-to-cement ratio, aggregate-to-cement ratio, RCA replacement ratio, parent concrete strength, maximum RCA size, maximum natural aggregate size, RCA bulk density, natural aggregate bulk density, RCA water absorption, natural aggregate water absorption, RCA Los Angeles abrasion and natural aggregate Los Angeles abrasion. Compressive strength was used as the sole output variable. Pearson correlation analysis showed that effective water-to-cement ratio had the strongest negative correlation with compressive strength (γ=−0.4835), followed by maximum RCA size (r=−0.3883) and aggregate-to-cement ratio (γ=−0.3390). An 80:20 training-testing strategy was adopted. RF achieved testing values of R²=0.8827, R²=0.7791, RMSE = 6.4922 MPa and MAE = 4.8381 MPa, whereas XGBoost achieved R²=0.9084, R²=0.8239, RMSE = 5.7973 MPa and MAE = 4.1539 MPa. XGBoost consequently demonstrated superior predictive performance across all principal evaluation criteria. The findings indicate that ensemble tree-based models can effectively capture the nonlinear relationships governing RAC compressive strength and provide a useful computational tool for preliminary structural-material assessment.