Filling in Missing Material Layer Information for Generating Wall and Slab Object Libraries from Drawing

Abstract: Automated generation of building information modeling objects from computer-aided design drawings require comprehensive information about material layers, including their functional roles—information typically absent from drawings. This study proposes a hybrid two-step methodology that combines a large language model and a machine learning algorithm to address it, focusing specifically on wall and slab objects. First, the method identifies the function of each material layer as a multi-class classification task, utilizing few-shot prompting with GPT-4o, achieving 99.8% macro F1-score. Second, it classifies whether material layer information pertains to walls or slabs using random forest model with FastText embedding, achieving 89.5% macro F1-score.