CC BY 4.0Shi, DachuanDachuanShiMeyer, OlgaOlgaMeyerFan, ZhiZhiFanWang, HaoHaoWangBauernhansl, ThomasThomasBauernhansl2025-09-012025-09-012026https://publica.fraunhofer.de/handle/publica/494757https://doi.org/10.24406/publica-523310.1016/j.rcim.2025.10310510.24406/publica-52332-s2.0-105012921927Reconfigurable manufacturing (RM) has emerged to support mass customization, which leads to frequent changes in production processes. RM necessitates the rapid reallocation of production resources to accommodate these evolving demands. To address this challenge, we propose a cognitive digital twin (CDT) system that integrates Asset Administration Shells (AAS) and large language models (LLMs) for adaptively matching between production processes and resource capabilities. Our approach centers on the structured representation of knowledge related to products, processes, and resources (PPR) using the AAS and leveraging this foundation for capability matching through the LLM. First, a methodology for developing interoperable AAS submodels (SM) is represented. Based on this, the SM templates of PPR are developed, serving as the knowledge base of the CDT. Next, we propose a capability matching mechanism using the LLM with chain-of-thought prompting. Finally, we design and implement an IT architecture that integrates an LLM-based retrieval-augmented generation system for executing capability matching alongside an AAS server for hosting AAS instances with dynamic values. The proposed CDT system enables the dynamic allocation of production resources to process steps, and is demonstrated and evaluated in a machining center use case. It effectively supports planning customized machining tasks through AAS-based knowledge representation and LLM-powered capability matching.entrueAsset administration shellCapability matchingCognitive digital twinLarge language modelReconfigurable manufacturingRetrieval-augmented generationjournal article