MACHINE LEARNING-BASED PREDICTION OF COMPRESSIVE STRENGTH OF RECYCLED AGGREGATE CONCRETE: A COMPARATIVE STUDY OF DIFFERENT ALGORITHMS

Abstract

This study presents a comprehensive empirical investigation comparing different machine learning algorithms for predicting the compressive strength of recycled aggregate concrete (RAC). A dataset comprising 385 experimental specimens was analyzed using five machine learning models: Random Forest (RF), Gradient Boosting Machine (GBM), Support Vector Machine (SVM), Artificial Neural Networks (ANN), and Multiple Linear Regression (MLR). The results revealed that the Gradient Boosting Machine achieved the highest predictive accuracy with an R² value of 0.947 and Mean Absolute Error (MAE) of 2.84 MPa, outperforming other algorithms. Random Forest demonstrated strong performance with R² = 0.931 and MAE = 3.21 MPa. Artificial Neural Networks exhibited competitive results with R² = 0.918 and MAE = 3.87 MPa. The input variables water-to-binder ratio, cement content, recycled aggregate replacement ratio, fiber content, and aggregate properties collectively explained 94.7% of the variance in compressive strength predictions. Statistical analysis confirmed significant relationships between material compositions and compressive strength (p < 0.001). Cross-validation analyses demonstrated GBM's superior generalization capability with minimal overfitting (validation R² = 0.943). These findings indicate that machine learning algorithms significantly enhance prediction accuracy compared to traditional regression methods, with an average improvement of 23.4% in predictive performance. The study establishes benchmarking protocols for algorithm selection in concrete strength prediction applications and provides practical guidance for engineers in material design optimization.

Keywords: Machine learning, recycled aggregate concrete, compressive strength prediction, gradient boosting, algorithm comparison