Chemistry-informed multi-objective mix design optimization of self-compacting concrete incorporating recycled aggregates

Abstract

One of the primary objectives of modern concrete advancements is to strike a delicate balance among factors such as improved mechanical performance, appropriate economic viability, and reduced carbon emissions. Given this backdrop, self-compacting concrete incorporating recycled aggregates (referred to herein as RASCC) has gained attention in the quest for sustainable construction materials. Nevertheless, concurrently attaining the above objectives is not an easy feat for this particular type of concrete. In this study, a machine learning model based on the XGBoost algorithm was developed using 368 sample data points to predict the compressive strength of RASCC. To enhance the model’s accuracy, the chemical composition of RASCC’s binding materials, rather than the quantities of constituent materials, was chosen as part of input parameters, giving rise to the term “chemistry-informed” for this model. The model’s interpretability was comprehensively examined using the SHAP library. Then, in conjunction with the explainable machine learning model, the NSGA-II algorithm was leveraged to establish a RASCC auxiliary design system, enabling triple-objective optimization (i.e., strength, cost, and carbon emissions). The findings indicated that the XGBoost-based model achieved superior accuracy in predicting RASCC’s strength as compared to an existing neural network-based model. Additionally, using compound contents as inputs imbued the model with chemical significance, thus further enhancing its accuracy and interpretability. In conclusion, this study presents a plausible and beneficial tool for the efficient, cost-effective, and low-carbon design of RASCC.