Investigation on compressive strength and splitting tensile strength of manufactured sand concrete: Machine learning prediction and experimental verification

Due to experimental constraints, traditional strength prediction models for manufactured sand concrete (MSC) exhibit significant limitations and have a narrow range of applicability. To overcome these limitations and enhance prediction accuracy, this paper, based on the widespread application of machine learning in concrete performance research, employed four algorithms: Support Vector Regression (SVR), Random Forest Regression (RFR), Gradient Boosting (GB), and Extreme Gradient Boosting (XGB) to develop highly accurate and strongly generalizable models for predicting the compressive strength (CS) and splitting tensile strength (STS) ofMSC. The predictive performance of each model was evaluated, the influence laws of various factors were analyzed, and finally, macroscopic experiments were conducted to validate the models' prediction accuracy and generalization ability. The development of these models and the analysis of the influence laws of various factors on the strength of MSC provide a valuable reference for the mix design. The conclusions are as follows: (1) The XGB model demonstrates the highest prediction accuracy and strongest generalization ability for both the CS and STS of MSC, achieving R-2 values of 0.98 and 0.94 on the testing set, respectively. (2) The CS and STS of MSC are primarily influenced by four factors: water-binder ratio (W/B), curing age (AGE), cement (C), and coarse aggregate (RI >= 0.08869, RI: relative importance), showing a positive correlation with AGE and C, and a negative correlation with W/B. As the fly ash content, stone powder content, maximum particle size of coarse aggregate, and sand ratio increase, the CS and STS initially increase and then decrease. Appropriately increasing the fineness modulus of manufactured sand is beneficial to both CS and STS, while the addition of slag is beneficial to CS and harmful to STS. (3) The prediction accuracy and generalization ability of the XGB model are verified through three sets of macroscopic experiments: C30, C35, and C40. The relative errors of each set's predictions are less than 8 %. (4) A graphical user interface is constructed based on the XGB model, enhancing its practicality.