Applications Of AI For Real-Time Mechanical Behavior Prediction In Deep Excavation Support Systems: A Critical Review

Progressive deformation in deep urban excavations can trigger catastrophic collapse before conventional monitoring and finite element method (FEM) back-analyses can effectively respond, primarily due to observational and computational latency. This paper presents a PRISMA-adapted review of 20 peer-reviewed studies (2021–2025) investigating artificial intelligence (AI) frameworks for real-time mechanical behavior prediction in deep excavation support systems. Two methodological streams are examined: data-driven models (LSTM, Transformers, ST-GAT) and physics-informed approaches (DeepONet, MS-SDL). By mathematically constraining deep learning within governing partial differential equations, these physics-informed models reduce the Root Mean Square Error (RMSE) of wall deflection predictions by 40–60% relative to standalone three-dimensional FEM baselines, sustaining predictive accuracy even when subjected to 50% sensor noise. However, field deployment remains heavily constrained by failure-state data scarcity, edge-computing latency, and algorithmic opacity. Advancing toward autonomous safety systems requires developing quantized networks for rapid edge inference, embedding interpretability through Physics-Informed Neural Networks (PINNs), and establishing strict regulatory frameworks to govern human-in-the-loop oversight on live construction sites.