Keqing Hu

Waste glass powder (WGP) is a promising recycled constituent for sustainable cementitious materials, yet its effectiveness depends strongly on the packing characteristics of the fine-aggregate skeleton. This study investigates the coupled effects of WGP replacement ratio and fine-aggregate gradation on the mechanical performance and microstructural evolution of cement mortar. Five aggregate gradations and four WGP replacement levels (0%, 8%, 10%, and 12% by mass) were evaluated through compressive and flexural strength tests at 7, 14, and 28 days, yielding 180 prism specimens in total. Scanning electron microscopy (SEM) and mercury intrusion porosimetry (MIP) were used to identify the governing microstructural mechanisms. Four regression models—linear regression (LR), support vector regression (SVR), random forest (RF), and extreme gradient boosting (XGBoost)—were compared for compressive strength prediction under identical data partitioning. XGBoost achieved the highest accuracy (R² = 0.9346, RMSE = 0.9860 MPa, MAE = 0.6943 MPa), outperforming RF (R² = 0.8896), LR (R² = 0.8845), and SVR (R² = 0.8612), and was subsequently coupled with particle swarm optimization (PSO) for constrained mixture design. Appropriate aggregate packing substantially improved WGP utilization by combining micro-filling, secondary pozzolanic reaction, pore refinement, and interfacial transition-zone densification. The best experimental combination (GCS10-M3) increased 28-day compressive and flexural strengths by 14.0% and 42.9%, respectively, relative to the reference mortar with the same gradation. The XGBoost-PSO framework identified a theoretical optimum containing 14.86% WGP, beyond the tested range, in a fines-rich bimodal aggregate gradation, with predicted 28-day compressive and flexural strengths of 41.10 MPa and 8.23 MPa, respectively. The results establish a mechanism-informed, data-driven route for designing resource-efficient mortar for building engineering applications.

With the rapid development of artificial intelligence (AI) technologies, machine learning (ML) has been widely applied in concrete material design, performance prediction, and intelligent structural engineering. Compared with traditional empirical approaches, ML can efficiently establish complex nonlinear relationships among concrete mix proportions, environmental factors, and performance indicators, thereby improving prediction efficiency, reducing experimental costs, and enabling multi-objective optimization of mix proportions. This paper systematically reviews the recent research progress of ML technologies in the field of concrete engineering, with particular emphasis on typical algorithms, including supervised learning, unsupervised learning, and reinforcement learning. Their applications in predicting workability, mechanical properties, durability performance, and mix proportion optimization are comprehensively summarized. In addition, recent advances in ML applications for crack detection and digital twin technologies are also discussed. Moreover, deep learning and computer vision (CV) technologies have significantly promoted the development of crack identification and structural health monitoring, whereas the integration of digital twin and Internet of Things (IoT) technologies has further expanded the application of ML in smart infrastructure. Finally, the current challenges associated with data quality, model interpretability, and engineering applications are summarized, and future research directions are discussed. Overall, by linking algorithm choice to specific concrete performance-prediction tasks, this review clarifies the conditions under which ML delivers reliable results and provides a structured reference for both researchers and practitioners. The comparative analysis indicates that ensemble tree-based models—particularly random forest and gradient-boosting variants such as XGBoost—together with well-tuned neural networks consistently achieve the highest predictive accuracy across most concrete properties, with reported test-set coefficients of determination commonly between 0.90 and 0.99, whereas limited data availability, inconsistent validation protocols, and restricted model interpretability remain the principal obstacles to engineering deployment.