The influence of spatial configuration on fire evacuation efficiency: a machine learning investigation

Prescriptive fire codes often operate under the assumption that meeting basic requirements, such as exit width, maximum travel distance, and the number of exits, is adequate for ensuring safe evacuation. This study examines the extent to which this assumption reflects evacuation performance by investigating the influence of architectural plan geometry using an integrated framework combining fire simulation, evacuation modeling, and machine learning (ML). Available Safe Egress Time (ASET) and Required Safe Egress Time (RSET) are assessed for residential and commercial occupancies and modeled using Gradient Boosting algorithms. The results indicate that architectural plan features are associated with nonlinear variations in both ASET and RSET. Parameters, including usable floor area, travel distances, corridor dimensions, exit capacity, and internal wall configuration, interact in a multi-parameter manner that is not fully represented by linear models. The higher predictive performance of Gradient Boosting relative to the tested baseline approaches supports the presence of complex relationships between plan geometry and evacuation metrics. Model interpretability analysis further indicates that the relative importance of architectural features differs between residential and commercial buildings. A divergence is observed between prescriptive compliance and performance-based evaluation, particularly in commercial occupancies. Several layouts classified as safe based on ASET/RSET criteria fail at least one of the implemented prescriptive requirements, while some prescriptive-compliant layouts are classified as unsafe under performance-based analysis. In contrast, residential occupancies show stronger agreement between the two approaches. The results suggest that evacuation safety is influenced by nonlinear interactions among architectural plan features and may not be fully characterized using a limited set of prescriptive criteria alone. Integrating ML algorithms with performance-based fire engineering provides a complementary quantitative basis for evaluating evacuation safety during the design process.