The observational method (OM) provides an effective framework for managing cost and risk associated with deep excavations; however, practical implementation is often hindered by limited site investigation data, complex subsurface heterogeneity, and discrepancies between design-stage predictions and field observations. To address these challenges, this study proposes an AI-empowered OM framework that integrates probabilistic uncertainty quantification and propagation at the design stage with real-time updating of numerical model predictions using field monitoring data during construction. Complex stratigraphy and spatially varying soil parameter fields are inferred from sparse data using machine learning to explicitly capture stratigraphic and parametric uncertainty. These geomaterial uncertainties, in conjunction with uncertainties associated with excavation support systems and construction processes, are propagated through random finite element analyses to yield distributions of high-resolution, stage-dependent excavation responses required for OM. During construction, Bayesian dictionary learning (BDL) progressively assimilates sparse, multi-source monitoring data to probabilistically update multi-type excavation responses. The proposed framework is demonstrated using a deep excavation case history. High-resolution geological models are reliably inferred from limited investigation data (e.g., 86.2% stratigraphic accuracy). Assimilation of early-stage monitoring data improves subsequent-stage predictions by 24.5–89.5% across response types within ten seconds, while reducing uncertainty and enhancing decision-making in practical OM implementation for deep excavations with complex ground conditions.
Tian et al. (2026) studied this question.