Plan2FEM: Can Large Multimodal Models Turn Plans into Finite-Element Graphs?

The process of interpreting architectural or structural plans in the form of finite element input is considered one of the main challenges in the field of structural engineering. The purpose of this paper is to investigate the capabilities of the recently proposed large language models (LLMs) in converting 2D plan images
directly into FE graphs in the form of nodes an members in JSON format. For this purpose, the concept of Plan2FEM is proposed, including task definition, strict schema definition, and validation repair process. The capabilities of the proposed models, including the popular GPT-5o, Claude, Qwen-VL and Llama Vision models, are evaluated using the proposed framework. The results of the proposed models indicate the better capabilities of the proprietary models in the filed of converting 2D plan images into FE graphs in the form of nodes and members in JSON format, especially in the case of cluttered plans with dense annotations.