AUTOMATION OF EARLY-STAGE BIM MODELING: A CASE STUDY INTEGRATING NATURAL LANGUAGE AND PYTHON

Abstract

This article aimed to develop an automation workflow for initial modeling in Building Information Modeling (BIM) by integrating natural language and programming language. The methodology consisted of a hybrid approach in which a Large Language Model (LLM) acts as a semantic interpreter to structure data in JSON format, while an IronPython algorithm operates as both an “auditor” and model generator through the Autodesk Revit API. This Python engine applies rigorous geometric validations, including restricted perimeter calculations, curve tolerance filters, and elevation constraint controls. The results demonstrated that delegating exclusive topological control to Artificial Intelligence leads to mathematical hallucinations and structural inconsistencies, such as pillar extrapolation across building floors. However, the intervention of the Python code mitigated 100% of these inconsistencies, enabling the autonomous and precise generation of structural skeletons, dynamic overhangs, and conceptual openings. It is concluded that the Hybrid Parametric Auditing method enables the reliable translation of modeling concepts expressed in natural language into consistent and geometrically sound parametric models. The approach demonstrated potential to optimize the early design stage, while overcoming geometric limitations inherent to generative artificial intelligence models, ensuring greater stability in the automated generation of BIM models.