From Data to Dialogue: Connecting Large Language Modelswith Digital Twins in Industrial Production

Modern industrial production systems challenge current digital twin (DT) approaches, which often rely on rigid rules, static dashboards, and task-specific views. These methods lack semantic reasoning, create high cognitive load, and limit flexible interaction, leaving many DT insights unused. Large language models (LLMs) address this gap by enabling semantic interpretation of heterogeneous data, natural language interaction, and domain knowledge integration. Acting as intermediaries, LLMs translate human queries into machine tasks and reframe DT outputs into human-understandable explanations. To support diverse monitoring, simulation, and control needs, we design a multi-agent architecture where specialized LLM agents cooperate, each focused on a functional area. This modular design allows task-specific reasoning and scalable adaptation. Using a design science approach, we derive requirements, build the architecture, and demonstrate it in a simulated scenario, showing benefits in explainability, adaptability, and human-centered interaction.