Enhancing Occupant Evacuation Simulation Using LLMs and Retrieval-Augmented Generation

Abstract

Effective evacuation simulations can be useful for assessing building safety design and optimizing emergency responses. However, configuring these simulations is often manual, time-consuming, and error-prone, especially with complex building geometries and diverse occupant characteristics. This paper introduces an automated workflow that integrates openBIM-based Occupant Movement Analysis data with fire safety regulations from the US and UK using Retrieval-Augmented Generation (RAG) methods and Large Language Models (LLMs). We benchmarked multiple parsing tools, with LlamaParse emerging as the most accurate for extracting text and tables from regulatory documents. We then tested eight RAG approaches with various LLMs across multiple question types and identified Knowledge Graph Enhanced RAG and Neo4j GraphRAG with GPT-4o as top performers in accuracy and consistency. These methods enabled on-demand interpretation of the US and UK regulatory documents for calculating occupant load from IFC-derived geometry data, generating a population input file for an Evacuationz simulation as a selected evacuation software in this study. Our evaluation confirms that Knowledge Graph Enhanced RAG excelled in complex reasoning while Neo4j GraphRAG offered higher stability. This automation enables efficient and reliable safety assessments, contributing to safer building design and emergency response planning.