Designing better hospitals: an adaptive simulation-optimization pipeline for layout planning

Hospital facility layout optimization combines difficult combinatorial structure with nonlinear operational dynamics. Classical formulations based on the Quadratic Assignment Problem are computationally tractable, but they rely on simplified costs that do not capture key hospital effects such as stochastic demand, urgency, resource contention, and temporal congestion. High-fidelity simulation captures these effects, but it does not expose the analytical structure required by standard optimization solvers. This thesis proposes a calibration-driven framework that bridges this gap by integrating a discrete-event simulator, an expert-defined evaluator, a mixed-integer linear programming optimizer, and a calibrator that learns feature weights linking the optimization objective to simulation outcomes. Calibration is performed with Bayesian optimization over black-box evaluations of the full pipeline. The framework is validated on synthetic scenarios and real operational data from Sant'Orsola Hospital (Bologna). In the real case study, the optimized layout reduces evaluated cost to approximately 79-80\% of the current baseline on both training and held-out periods, indicating robust improvements beyond the calibration window. Across synthetic experiments, the method consistently outperforms naive weighting strategies and shows strong instance-dependence of optimal weights, motivating an explicit per-instance calibration process.