Large language models as a decision-making core for semantic capability check in asset administration shell-based intelligent manufacturing

Abstract Flexible manufacturing and lot-size-one production require deciding, at run time, whether an available resource can perform a required operation. The IDTA 02020 Capability Description submodel standardizes how asset capabilities are represented in the Asset Administration Shell (AAS), but it provides no mechanism to reason over and compare them. This article investigates the use of a Large Language Model (LLM) as the decision-making core of a capability-matching procedure that operates directly over this standardized submodel, without shared ontologies. The approach is evaluated under the variability with which the same capability can be described in the AAS, across sixteen test cases spanning four abstraction dimensions and two degrees of semantic formalism. Process-level semantic alignment, anchored in the DIN 8580 taxonomy, was correct in all cases and indifferent to the degree of formalism of the description, whereas requirement verification under a conservative aggregation policy rejected every positive case while producing no false positives. The analysis traces this systematic rejection to a single, correctable cause: a boundary that the standardized submodel does not disambiguate — between properties that constitute requirements to be proven and properties that merely describe the product. The contribution is therefore less the aggregate verdict than the localization of the layer at which LLM-based matching over the standardized submodel is reliable and the precise mechanism by which it fails, together with the modeling and prompting strategies that follow from this diagnosis.