By integrating modern methods of machine learning and building evacuation modeling, it is possible to test new approaches for estimating the development of emergency crisis events to manage them more effectively. So far, the use of evacuation simulation models has been primarily in the design phase of construction projects, where such models enable the analysis of different building design variants in terms of evacuation efficiency. A current challenge in this field is the application of simulation outputs during emergencies in the operational phase, for which these models are not primarily intended. By replacing the simulation computation process with real-time estimations of their outputs, it is possible to obtain valuable information that reflects current building conditions (e.g., the number and location of occupants) to support effective decisionmaking by management or emergency services during evacuation and to mitigate the negative impacts of the emergency. The dissertation tests the replacement of evacuation model simulations with realtime estimates provided by machine learning regression models trained on synthetic simulation datasets (so-called surrogate evacuation models). The testing was carried out on basic tasks and subsequently on complex evacuation scenarios in buildings and road tunnels. It was demonstrated that machine learning models can estimate the time required for evacuation with sufficient accuracy in scenarios with congestion, and—with an extended training dataset—even in highly variable scenarios. Among all the tested models, artificial neural networks proved to be the most accurate across all scenarios. The results revealed the potential of the proposed approach for possible real-world application.