Ontology-Based Grounding and Feedback for Reliable LLM-Based Asset Administration Shell Generation

IEEE International Conference on Emerging Technologies and Factory Automation (ETFA), pp. 1–8

Abstract

Creating and maintaining high-quality, consistent and interoperable Asset Administration Shells (AAS) requires skilled engineers, with expertise in the AAS metamodel, relevant submodel templates, and the application domain, which limits broader adoption of AAS-based digital twins in manufacturing. This paper presents a modular framework that combines Large Language Model (LLM)-based AAS generation with ontology-driven validation and human-in-the-loop correction. Central to the framework is the AAS Resource Structure Ontology (ARSO), which extends the AAS ontology with resource-specific submodel structures and cross-submodel dependency rules. ARSO is used both to guide generation through prompt context and to derive SHACL shapes for formal post-generation validation. The framework follows a three-layer architecture: a generation layer produces a compact AAS profile, a transformation layer expands it into a complete AAS instance, and a validation layer projects the AAS to RDF and checks it against ARSO and AAS constraints. When violations are detected, structured validation reports are returned to the LLM for iterative correction. Remaining issues can be inspected and resolved in an interactive editor with live validation feedback. The framework is evaluated on two synthetic industrial use-case assets using datasheets and communication specifications, across four LLMs from Anthropic and Google. The results show that ontology-based validation enables most generated AAS instances to reach SHACL conformance within a limited retry budget of three generation attempts. Larger models, such as Claude Sonnet~4.6 and Gemini~3.1~Pro, follow the required output format more reliably, make better use of validation feedback, and reduce the manual effort needed to finalise the AAS. The results indicate that combining LLMs with ontology-based grounding, formal validation, and interactive human verification is a promising approach for scalable AAS generation.

Authors 5

1. Martin B. Jensen

   Aalborg University

   Affiliation as printed

   Aalborg University

2. Tristan Schwörer

   Aalborg University

   Affiliation as printed

   Aalborg University

3. Sazvan Saeed Aachen Chair of Intelligence in Quality Sensing

   RWTH Aachen University

   Affiliation as printed

   Chair of Intelligence in Quality Sensing, WZL RWTH Aachen

4. Zakka Ugih Rizqi

   Aalborg University

   Affiliation as printed

   Aalborg University

5. Casper; id_orcid 0000-0001-7831-311X Schou

   Aalborg University

   Affiliation as printed

   Aalborg University