Development of chemistry-informed interpretable model for predicting compressive strength of recycled aggregate concrete containing supplementary cementitious materials

Abstract

Accurate prediction of compressive strength of recycled aggregate concrete (RAC) has great value for promoting the sustainable development of concrete field. The study aims to propose a chemistry-informed interpretable model for predicting compressive strength of RAC containing supplementary cementitious materials. First, this study collected the compressive strength records of RAC. Then, the unsupervised learning algorithm isolation forest was utilized to perform data cleaning, and 1600 samples were obtained to build the dataset. And the topological network constraint was introduced to represent the reactivity of cementitious systems. The five machine learning (ML) algorithms were employed to predict compressive strength of RAC, and the prediction accuracy of above models were compared. In addition, SHapley Additive exPlanations (SHAP) method was integrated with the ML model to shed more light on the outcome of ML model. Results showed that the Categorical Boosting (CatBoost) outperformed other ML models on the test set, and the coefficient of determination (R2), root mean square error (RMSE), and mean absolute error (MAE) were 0.961, 2.650 MPa, and 1.636 MPa. The introduction of chemical information enhanced the predictive accuracy of model, and the R2 increased from 0.935 to 0.961, the RMSE and MAE decreased by 22.7% and 15.1%, respectively. The SHAP values of curing age, natural coarse aggregate replacement ratio and water-binder ratio were higher than those other input variables, indicating that the three features were essential for predictive results of model. And the curing age played a positive role in the strength development of RAC, and the increase of natural coarse aggregate replacement ratio and water-binder ratio caused the loss of compressive strength. The influence trends of features on the outputs were in accord with the engineering experience, which proved the reliability of predictions. The contribution and importance of features on the predictions provide a theoretical guidance to aid the concrete mixture design for engineers to obtain the desired performance.