Large-scale construction projects are characterised by uncertainty, multidisciplinary dependencies, and frequent cost and schedule overruns. Structural engineering and project planning increasingly use probabilistic methods, but within separate workflows, so uncertainty information is often lost when results are transferred manually between disciplines. As a result, projects experience fragmented workflows, which negatively impact project performance. To address this, the thesis develops and evaluates an AI-supported decision framework that integrates probabilistic structural reliability analysis with probabilistic project planning.

The framework consists of three components. The first one is a Probabilistic Surrogate Module (PSM), which combines principal component analysis and Gaussian process regression to approximate the structural response of an immersed-tunnel cross-section, achieving a coefficient of determination of 0.991 at approximately three orders of magnitude lower computational cost compared to the reference model. Combined with Monte Carlo simulation, it estimates the probability of failure and reliability index while accounting for aleatoric and epistemic uncertainty. Secondly, a Probabilistic Planning Module (PPM) is integrated into the framework, which is based on the Mitigation Controller software. This module propagates uncertainty in durations, costs, risks, and mitigation measures into the project schedule. A coupling mechanism is used to translate structural reliability information into the occurrence probability of a planning risk. Lastly, a human-controlled AI agent was developed that functions as an orchestration layer between the mathematical modules and the human decision-maker. All three of these components are implemented into an interactive dashboard with which the human decision-maker can freely interact.

The framework was applied to the Fehmarnbelt Tunnel through four validation scenarios and evaluated in sessions with six practitioners. The results demonstrate functional feasibility and professional plausibility. The coupling mechanism, however, is not yet empirically calibrated, and measurable improvements in decision quality have not yet been established.

The main contribution lies in system-level integration of novel methods within an uncertainty-aware, human-controlled decision environment.
... Read more