A Deep Learning-Based Framework for Automated Section Line Generation in BIM Models

Abstract

Despite increasing Building Information Modeling (BIM) adoption in the Architecture, Engineering, and Construction (AEC) industry, 2D drawings remain necessary as construction project deliverables. Previous studies on BIM-based 2D drawing automation have mainly focused on fixed auxiliary elements, such as dimension lines and grid lines, whose locations can be determined using predefined rules. However, section lines have flexible locations because their placement depends on the information designers intend to communicate through section views. In addition, AI-based methods for predicting such locations generally require manually labeled data, resulting in substantial preprocessing effort. To address these limitations, this study proposes a deep learning-based framework with two models for different stages. The AutoLabel model automatically generates labeled data from raw 2D drawings, whereas the AutoSection model learns section-line placement patterns from labeled data and generates section lines. A Dynamo-based parametric algorithm then applies the generated section lines to BIM models to create corresponding section views. The framework was evaluated using 199 2D drawings collected from residential and commercial buildings, with the final dataset comprising 3,184 images after data augmentation. AutoLabel generated labeled data with an average structural similarity index measure (SSIM) of 92.7% within 1.87 s per drawing. AutoSection, consisting of Pix2Pix-GAN and post-processing, achieved an average SSIM of 95.39% for section-line generation. The proposed framework extends BIM-based 2D drawing automation from rule-based placement of fixed elements to learning-based placement of flexible section lines. It also reduces manual effort in labeled-data generation and has the potential to improve productivity in 2D drawing generation.