Doc2DT: An LLM-driven knowledge graph abstraction framework for digital twin generation from unstructured engineering documents

Engineering documents contain substantial knowledge about the structure, behavior, and operation of physical systems, yet this knowledge is typically expressed in heterogeneous human-readable forms that are difficult to directly reuse in computational engineering workflows. As a result, the development of digital twins often relies on domain-specific modeling procedures and extensive manual engineering effort, limiting the reuse of generation methodologies across different systems. To address this challenge, this paper proposes Doc2DT, an LLM-driven knowledge graph abstraction framework that transforms unstructured engineering documents into reusable digital twin representations through a unified abstraction process. The framework converts document-level engineering information into structured digital representations and reformulates them as graph-based abstract models that explicitly capture engineering knowledge in a machine-interpretable form. The resulting abstraction serves as the foundation for digital twin generation and system integration. The proposed framework is evaluated through industrial robot case studies in which engineering documents are transformed into validated digital representations and deployed as operational digital twins. Additional robot casesfurther demonstrate that the same abstraction methodology can be retained despite differences in geometry, document structure, and local ambiguity. The results support the effectiveness, generality, and scalability of the proposed abstraction-centered approach within the articulated robot domain and indicate a broader pathway toward reusable digital twin generation from heterogeneous engineering documents.