Building Information Modelling (BIM) data provides an abundant source with hierarchical and detailed info rmation on architectural elements. Nevertheless, transforming BIM data into an understandable format for AI to learn and generate controllable and detailed three-dimensional (3D) models remains a si gnificant research challenge. This paper explores an encoding approach for conver ting BIM data into graph-structured data for AI to learn 3D models, which we define as Graph-BIM encoding. We employ the graph reconstruction capabilities of a Variational Graph Autoencoder (VGAE) for the unsupervised learning of BIM data to identify a suitable encoding method. VGAE's graph generation capabilities also reason for spatial layouts. Results demonstrate that VGAE can reconstruct BIM 3D models with high accuracy, and can reason the en tire spatial layout from partial layout information detailed with architectural components.
| Authors | Jiadong Liang and Ximing Zhong and Immanuel Koh |
|---|---|
| Year | 2024 |
| Venue | Proceedings of the International Conference on Computer-Aided Architectural Design Research in Asia |
| DOI | 10.52842/CONF.CAADRIA.2024.1.221 |
| Source Database | Seed (CumInCAD) |
| Bridge-to-GNN | Category A |
| GNN Architecture | Graph GAN (heuristic — verify) |
| Graph Encoding | BIM Element Topology Graph (inferred from title) |
| AEC Task | Floorplan / Layout Generation (inferred from title) |
| Cohort | Mature Applications (2024–2026) |
| Implementation Framework | Revit / Dynamo (heuristic — verify) |
|---|---|
| Primary Metric | Accuracy (heuristic — verify) |
| Key Finding | incorporating room function connections as special attributes within our encoding framework (heuristic — verify) |
| Quality Assessment | 1/6 · rigor: Low · Baseline: ✓ (heuristic — verify) |
| GNN architectures detected | Graph GAN · Graph VAE / VGAE · Knowledge Graph Embedding (TransE/RotatE/DistMult) · GCN |
|---|---|
| Frameworks / libraries | Revit / Dynamo |
| Quality (heuristic, 6-flag) | 0/6 · rigor: Low Code/Data: —Benchmark: —Baseline: —CV/Split: —Ablation: —Reproducible: — |
Part of the GML/GNN in AEC Systematic Review (PRISMA 2020) — 112 papers, 2020–2026