Optimizing compressive strength of quaternary-blended cement concrete through ensemble-instance-based machine learning

Abstract

Optimizing the composition of concrete, particularly when incorporating ordinary Portland cement (OPC) along with supplementary materials such as ground granulated blast furnace slag (BF), fly ash (FA), and silica fume (SF), poses challenges due to the intricate nonlinearity inherent in concrete properties. Addressing this challenge has urged growing interest in employing Machine Learning (ML) techniques for precise property assessment. In this investigation, various predictive models, encompassing eXtra Gradient Boosting (XGB), random forests (RF), and K-Nearest Neighbor (KNN) algorithms were formulated to predict the compressive strength (CS) of ternary-blended concrete incorporating OPC, BF, FA, and SF. Through comprehensive analysis of an extensive dataset comprising 810 distinct records, the study identifies optimal concrete blends, providing invaluable insights for practical applications. Evaluation of ML models underscores the superior performance of XGB, with a coefficient of multiple determination (R2) values of 0.923 and 0.997 for training and testing datasets, respectively, indicative of exceptional predictive accuracy. Additionally, SHAP values underscore the importance of input age, water–binder ratio, and OPC content. This research enhances understanding of ternary-blended concrete characteristics, offering practical implications for construction practices and laying a foundation for future research endeavors to refine ML models, explore attribute interactions, and optimize concrete constituent proportions for broader civil engineering applications.