A systematic review and thematic analysis of physics-informed machine learning in construction and infrastructure systems

This study conducts a systematic review and thematic analysis to explore how physics-informed machine learning (PIML) can enhance reliability in infrastructure assessment and construction management. Previous studies are classified based on state estimation, digital twin updating, performance prediction, deterioration prediction, multi-hazard response analysis, and maintenance and inspection decision support. Embedding governing physical laws in the learning process ensures that models adhere to mechanical principles. The analysis shows that PIML can enhance prediction stability, decrease data needs, and allow for the identification of concealed structural and construction parameters through inverse methods. Importantly, integrating mechanics into data-driven models transforms machine learning from a curve-fitting tool into a verification-compatible analysis method. Constraining predictions with physical laws reduces unrealistic outputs; enables extrapolation to unobserved hazard scenarios such as earthquakes, winds, and floods; and facilitates estimation of deterioration states needed for risk-informed maintenance planning. The reviewed literature shows that these capabilities enable operational digital twins that continuously update infrastructure conditions using real-time monitoring data. However, there are persistent challenges, including computational expenses, inadequate uncertainty quantification, limited multi-physics integration, and the lack of benchmark datasets and regulatory approval. This review concludes that PIML offers a pathway to next-generation artificial intelligence in construction and infrastructure systems by bridging monitoring data and engineering mechanics. As a reliability engine within digital twin frameworks, PIML supports predictive maintenance, performance-based design, and resilience-oriented management of existing and future infrastructure systems.