Enhancing Occupant Evacuation Simulation Using LLMs and Retrieval-Augmented Generation

Abstract:

Background: Evacuation simulations are vital for assessing the safety of building occupants and identifying vulnerabilities in fire safety design. Traditionally, configuring these simulations involves manually inputting complex parameters like building geometries, occupant characteristics, and hazard scenarios. This process is time-consuming, labor-intensive, and prone to human errors. Recent international efforts have focused on developing an open building information modeling (openBIM) digital workflow for occupant movement analysis (OMA) to address these challenges and prevent costly and dangerous design errors, as highlighted by the inquiry into the UK Grenfell Tower disaster. However, the current openBIM OMA standard has only focused on the physical aspects of modeling occupancy space. This research introduces an innovative method to further streamline the setup and configuration of evacuation analysis through an AI-enhanced design process.

Objective: his study investigates the feasibility of leveraging Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG, which combines language generation with information retrieval) to support engineers with the selection of input parameters for evacuation models to enhance accuracy and efficiency in the simulation configuration process.

Methods: Leveraging the latest openBIM standard for occupant movement analysis, we developed a framework that integrates LLMs with RAG to generate simulation inputs based on natural language descriptions of building layouts, occupant distributions, and hazard scenarios. A comprehensive knowledge base was constructed, encompassing domain-specific data such as OMA standards. The Evacuationz evacuation model (a tool for simulating occupant movement) was used to test these configurations. Comparative analyses were conducted between a codified approach, such as using default evacuation times for a typical hotel building, and a performance-based approach, where specific evacuation times were determined dynamically based on building usage and occupant behavior.

Results: Preliminary results indicate that the use of LLMs combined with RAG significantly improves the accuracy and efficiency of simulation configurations compared to traditional manual methods and standard LLM approaches without RAG enhancement. User studies demonstrated increased satisfaction due to reduced configuration time and minimized errors.

Conclusion: Integrating LLMs and RAG into the configuration process of crowd evacuation simulations enhances efficiency, accuracy, and accessibility. This approach holds promise for more precise safety assessments and informed decision-making in fire safety and civil engineering, complementing existing efforts to develop openBIM-based digital workflows.

Keywords: Large Language Models, Retrieval-Augmented Generation, Evacuationz, Building Evacuation Simulations, Pedestrian Models, Simulation Configuration, Safety.