Interpretable machine learning model for compressive strength prediction of self-compacting concrete with recycled concrete aggregates and SCMs
Qinglong Miao, Zhiwei Gao, Kangjian Zhu, Zhanggen Guo, Qiansen Sun, Ling Zhou
Abstract
The composite comprising waste concrete and industrial by-products for producing recycled aggregate self-compacting concrete (RA-SCC) mitigates natural resource depletion and reduces greenhouse gas emissions, thus offering significant environmental benefits. This study aims to predict the compressive strength of RA-SCC by using five interpretable machine learning (ML) methods including M5P model tree, Support Vector Regression (SVR), Multilayer Perceptron (MLP), Random Forest (RF) and Extreme Gradient Boosting (XGBoost). To develop the ML models, a dataset containing 454 samples was constructed, encompassing experimental data with varying material dosages, curing ages, and recycled aggregate properties. The hyperparameters were optimized using Bayesian optimization and 10-fold cross-validation. Additionally, the SHapley Additive exPlanations (SHAP) method was employed to assess the importance and contributions of the input parameters affecting compressive strength. The results indicated that the XGBoost model achieved the highest accuracy among the tested models, with R 2 = 0.909, RMSE = 4.99 MPa, and MAE = 3.54 MPa. SHAP analysis revealed that cement content and age were the most influential factors positively affecting compressive strength predictions, while fly ash and recycled concrete aggregate content exhibited the highest negative effects. The ML models proposed in this study provide a systematic evaluation of the compressive strength predictions for RA-SCC and offer valuable insights for concrete mix design and quality control.