Enabling Non-Simulation Experts to Adapt Digital Twins during Production System Redesign using Agentic AI

Simulation-based digital twins are a core foundation for the efficient and holistic optimization of material-flow systems across the lifecycle - from design and planning to control. However, constructing digital representations of supply chains and production/logistics systems is labor-intensive and largely manual. Current research seeks to shift from manual expert modeling to auto-generated digital twins derived from event-log data using process mining and machine learning. Yet, structural adaptation of these models during (re)design typically still requires intervention by simulation experts. This paper presents an approach that enables structural adaptation of digital-twin models during redesign via an LLM-based design assistant, which performs modeling through natural-language interaction with non-expert users. The assistant employs an agentic, retrieval-augmented generation (RAG) workflow and relies exclusively on publicly available information from the open-source digital-twin platform OpenFactoryTwin (OFacT) to execute adaptations. The proposed framework is evaluated in an empirical study based on data from an IoT-Factory - a laboratory assembly environment.