ACDT: A Distributed Agentic Cognitive Digital Twin Architecture

Abstract

Digital Twins are increasingly adopted to support monitoring, simulation, and decision-making in Cyber-Physical Systems and intelligent IoT environments. However, conventional Digital Twins provide limited cognitive capabilities, while directly employing Large Language Models (LLMs) as controllers may compromise controllability and the physical consistency of generated responses. This paper proposes ACDT, a Distributed Agentic Cognitive Digital Twin architecture in which an LLM acts as an Orchestrator Agent that coordinates specialized Digital Twin agents rather than directly controlling the system. The proposed architecture combines explicit knowledge partitioning, distributed evidence gathering, and Bayesian-supported explainability to integrate runtime data, simulation services, SysML-based engineering knowledge, and diagnostic reasoning within a unified cognitive framework. The architecture is validated through the BIOPURE case study using a MATLAB-based Digital Twin and a locally deployed Llama 3.2 3B model. A claim-based evaluation demonstrates that combining runtime and engineering knowledge significantly improves the correctness and completeness of the generated responses while reducing unsupported claims. The results show that the proposed architecture provides an effective and explainable foundation for trustworthy Cognitive Digital Twins supporting complex engineering processes.