Ensuring safe and efficient building evacuations is a central challenge in safety engineering and architectural design. Microscopic evacuation simulators such as crowd:it provide high-fidelity insights into crowd behavior, yet their computational cost makes them impractical for large-scale parameter studies or iterative design workflows. This thesis investigates surrogate modeling as a means to accelerate evacuation analysis by predicting macroscopic evacuation outcomes, specifically total evacuation times, at a fraction of the simulation cost.
We applied the SWIM algorithm (Sampling Where It Matters) to train neural surrogates on two complementary datasets. Scenario 1, a synthetic room-based setup, established a controlled baseline and confirmed that the method works in a simplified environment. Scenario 2, based on the regulation-derived Gd99 vertical-evacuation dataset, demonstrated applicability to real-world benchmarks and additionally served as the basis for evaluating sampling strategies. These included Quasi-Monte Carlo (QMC), Randomized QMC (RQMC), entropy-based, sparse-grid, and clustering approaches, applied to determine how few simulations are needed to train reliable surrogates.
The results show that SWIM-based surrogates can reproduce macroscopic evacuation outcomes with high accuracy while reducing simulation demand by more than an order of magnitude. Sampling strategy proved critical, with RQMC delivering the strongest generalization and achieving predictive accuracies above R2 = 0.95 using only around a dozen simulations, while other methods required substantially larger sample sets for comparable performance.
Overall, this thesis confirms that surrogate modeling with SWIM is a promising tool for both research and applied safety practice. By integrating such approaches into existing simulation platforms, evacuation analysis can be performed rapidly, supporting both regulatory compliance and innovative building design.