AASia: An Intelligent Framework for Generating Asset Administration Shells from Natural Language

This dataset supports the validation of the AASia framework, an LLM-based pipeline for generating Asset Administration Shell (AAS) instances from unstructured natural language descriptions of industrial assets. The dataset is organized into three main parts.

AASIA-Manual-AAS-Modeling-Study contains the artifacts from a comparative study of manual AAS modeling. Three industrial test cases — an electric motor (TC1), a centrifugal pump (TC2), and a filling and capping machine (TC3) — were independently modeled by three graduate-level engineering participants using the AASX Package Explorer. The resulting AAS artifacts are provided in .aasx and JSON formats, together with participant evaluation spreadsheets reporting estimated modeling time, perceived effort, and qualitative observations.

AASIA-Automated-AAS-Modeling-Study contains the AAS artifacts generated by AASia for the same three test cases, enabling direct structural comparison with the manually created artifacts. For TC3, two executions are included: one using the original input description and one using a revised version with increased explicitness regarding quantitative ranges and units.

AASIA-Robustness-Evaluation contains the artifacts from a broader robustness evaluation of the AASia pipeline under its stabilized configuration. A total of 26 executions were performed across 9 distinct industrial asset categories — storage tanks, pressure sensors, homogenizers, spray dryers, air compressors, control valves, belt conveyors, plate pasteurizers, and mixing reactors — using input descriptions ranging from 15 to 150 words in both Spanish and English. Two LLM backends were evaluated: the Groq API with meta-llama/llama-4-scout-17b-16e-instruct and a local LM Studio deployment with meta-llama/llama-3.1-8b-instruct. All 26 executions produced structurally valid AAS Environment JSON documents and AASX packages. Each execution folder contains the input description, the intermediate structured interpretation, the generated AAS JSON, the exported AASX package, and execution logs. The file dataset_index.csv provides a structured summary of all executions, including asset type, input language, LLM configuration, word count, pipeline status, and number of submodels and properties generated. Failed execution attempts are also included for documentation purposes, with observer notes describing the observed issue and resolution.

The complete dataset is intended to support reproducibility of the reported analyses, facilitate independent inspection of the generated artifacts, and enable further research on automated AAS generation approaches.