Toward Explainable, Probabilistic Risk Control in Metro Tunnel Megaprojects: An XAI-Enabled Bayesian Survival–Cost Framework

Abstract Mechanized metro tunneling involves tightly interdependent geotechnical, logistical, and governance risks the evolving interactions of which can trigger stoppages, excessive settlement, material-supply disruptions, and permit-related delays. This study developed an integrated and explainable risk-to-decision framework that closes the loop between prediction, interpretation, and action. An explainable artificial intelligence (XAI) layer, combining global and local attribution analysis with counterfactual reasoning, reveals how measurable operational and project-control variables influence both hazard dynamics and downstream performance. Building on these outputs, a conditional value at risk (CvaR)-based portfolio optimization module uses posterior scenario draws and lever-specific counterfactual shifts to identify mitigation bundles that are feasible under budgetary, operational, and eligibility constraints, including interventions such as targeted pregrouting, buffer inventory, appointment metering, and permit front-loading. The framework was validated on 41 mechanized tunnel drives across 11 US metropolitan areas between 2013 and 2025, covering 153.4 km and 1,680 chainage buckets. Relative to Cox-based benchmarks, the survival module improved cause-specific predictive performance, reaching a C-index of as much as 0.88 and reducing the integrated Brier score by 0.026–0.036. The conformally calibrated downstream heads achieved near-nominal interval reliability, with 90% empirical coverage of 0.90 for cost and 0.89 for schedule. These findings translate directly into governance-ready operational triggers, such as maintaining yard utilization below the mid-80% range and stabilizing face-control variability. The optimized mitigation portfolios yielded substantial tail-risk reduction, including an overall 24% decrease in CVaR 0.90 and case-specific reductions ranging from 10% to 26%, thereby supporting contingency allocation, change-order justification, and lender-facing risk reporting through transparent and auditable uncertainty quantification. The proposed framework is among the first in mechanized metro tunneling to integrate hierarchical competing-risks survival modeling, calibrated downstream interval prediction, explainability explicitly linked to field levers, and CVaR-based constrained portfolio selection within a single coherent and auditable workflow.