Facilitating the application of recycled concrete UHPC: compressive strength prediction and statistical association analysis based on explainable machine learning
Recycled concrete ultra-high performance concrete (RC-UHPC) prepared from recycled concrete sand (RCS) and recycled concrete powder (RCP) represents an important pathway for achieving the high-value utilization of construction and demolition waste. However, influenced by the material properties and compositional complexity, the application of RC-UHPC requires extensive performance optimization trials, a process that is not only inefficient but also accompanied by high costs. To address this challenge, this study introduces machine learning (ML) methods to improve the design efficiency of RC-UHPC, reduce reliance on conventional experimental methods, and facilitate its engineering application. A feature system comprising 11 key input variables, including cementitious materials, admixtures, and RCS with different particle sizes, was established. The performance of various models in predicting the 28-day compressive strength (CS28) of RC-UHPC was systematically evaluated, and the main influencing factors were identified through feature analysis. The results indicate that k-nearest neighbors, random forest, and gradient boosting regression trees are the optimal ML models, with mean R2 values of 0.974 ± 0.014, 0.954 ± 0.008, and 0.971 ± 0.007, respectively, under different random seeds. Feature analysis indicates that steel fibers and fly ash have a significant positive effect on CS28. In contrast, RCS and RCP exhibit a deteriorating effect, and no significant difference is observed in the effect of RCS with different particle sizes (0–1.18 mm and 0–2.36 mm) on CS28. This study analyzes the statistical associations between RC-UHPC performance and its influencing factors from a data-driven perspective. The results can provide scientific guidance for mix design and performance regulation of RC-UHPC in engineering applications.