Generative AI-Powered Digital Twins for Crowd Management in Large-Scale Events

Résumé fourni par la source

Managing safety and operational efficiency in large-scale events requires decision-support tools capable of representing complex crowd dynamics while enabling rapid and evidence-based operational assessment. This paper presents a Generative AI-driven simulation-enabled digital twin prototype that integrates an agent-based crowd simulation framework, an API-based execution pipeline, and a Large Language Model (LLM)-driven conversational interface within a unified architecture. The proposed framework enables the dynamic configuration, execution, and analysis of crowd scenarios under different operational conditions, including high-demand situations and emergency evacuation contexts. Experimental results show that the system can reproduce nonlinear crowd dynamics, identify congestion patterns, and assess evacuation performance. While evaluated under TRL-4 conditions, these results demonstrate the framework’s architectural potential to provide actionable insights for planning and safety evaluation once empirically calibrated with real-world data. A central contribution of this work is the introduction of an API-based execution paradigm that exposes the complete simulation lifecycle, including configuration, validation, execution, and output retrieval, through programmatic interfaces. This design supports reproducible, modular, and scalable what-if analysis. In addition, the integration of an LLM-based conversational interface allows non-technical users to interact with complex simulation models through natural language, improving accessibility without compromising execution control. The framework is validated through a TRL-4 prototype, demonstrating stable performance and reliable interaction behavior. Scalability is strictly confirmed within the evaluated hardware configuration, model abstraction level, and tested agent scale (up to 60,000 agents), providing a foundation for localized event management. Overall, the proposed system serves as a simulation-enabled digital twin prototype, demonstrating how models can transition from static analytical representations toward executable, interactive, and user-centered platforms, laying the necessary architectural groundwork for future operational decision support in complex urban environments.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.