The use of neural networks to retrieve relevant images has become mainstream. However, retrieving images that contain speci fic spatial relationships remains a challenging task. Images alone are not suf ficient to fully describe spatial and topological relationships, which are usually better represented as a graph made up of nodes and edges. This paper describes the development of a graph-based computational tool for retrieving architectural precedents that closely match the relationship between a building and its surrounding ground as detected in a designer ’s project. The tool, titled Building Ground Relationship (BGR), stems from a research project into Graph Machine Learning (GML) that used Deep Graph Convolutional Neural Networks (DGCNNs) to classify building and ground relationships. The neural network was trained using a large synthetic dataset of graphs and optimized through the fine-tuning of its hyperparameters.
| Authors | Alymani, Abdulrahman and Jabi, Wassim |
|---|---|
| Year | 2025 |
| Venue | International Journal of Architectural Computing |
| DOI | 10.1177/14780771241260853 |
| Source Database | EBSCO (Applied Science & Technology) |
| Bridge-to-GNN | A + B (extracted) |
| GNN Architecture | DGCNN (3 conv layers × 32 neurons, SortPooling, 1D conv, MaxPooling, dense + SoftMax) and DGL with SAGEConv (3 hidden layers × 32 neurons, Adam, MaxPooling) (extracted) |
| Graph Encoding | Homogeneous, undirected, 3D topological (extracted) |
| AEC Task | Graph Classification (5 classes: separation, separation with plinth, adherence, adherence with plinth, interlock) (extracted) |
| Cohort | Mature Applications (2024–2026) |
| Category Detail | Core GNN/GML (DGCNN + DGL classification) with strong Topology/Graph Data Infrastructure (Topologic-derived dual graphs from 3D models) (extracted) |
|---|---|
| AEC Sub-Domain / Task | Architectural Design / Building–Ground Relationship Classification and Precedent Retrieval (extracted) |
| Dataset | Custom synthetic BGR dataset (extracted) |
| Data Type (Raw Input) | 3D topological dual graphs generated from parametric models via Rhino Grasshopper + Topologic (extracted) |
| Dataset Size | 2,136 graphs; 171,232 nodes; avg 80 vertices/graph (min 20, max 258). Flat: 240, Sloped: 684, Level: 1,242 iterations (extracted) |
| Node Definition | Architectural and spatial elements from non-manifold topology of 3D conceptual model (extracted) |
| Edge Definition | Topological connections and spatial relationships between building elements and ground (extracted) |
| Graph Type (detail) | Homogeneous, undirected, 3D topological (extracted) |
| Additional Methods | DGL used as surrogate/verification model (extracted) |
| Implementation Framework | PyTorch (DGCNN); DGL in Jupyter Notebook (extracted) |
| Key Hyperparameters | DGCNN: 100 epochs, LR 1e-3, batch 1, 70/30 split. DGL: 100 epochs, LR 0.0001, batch 1, 80/20 split (extracted) |
| Primary Metric | Accuracy (extracted) |
| Primary Value | DGCNN: 99.69% (loss 0.006); DGL: 99.8% (loss 0.14) (extracted) |
| Secondary Metrics | Confusion matrix: DGCNN 638/640 correct (2 misclassified Class 4→Class 2); DGL 639/640 correct (1 misclassified) (extracted) |
| Baselines Compared | DGCNN vs DGL surrogate model (internal cross-validation, not external baselines) (extracted) |
| Key Finding | Both DGCNN and DGL achieve >99.5% accuracy on topological graph classification of building–ground relationships; DGL slightly outperforms with fewer misclassifications (extracted) |
| Reported Limitations | Switching between software platforms presented challenges. Synthetic dataset limited to 2,136 elements, potentially constraining ML model generalisability. (extracted) |
| Proposed Future Work | Expand dataset with semantic information (apertures, environmental analysis, setbacks). Apply to real 3D BIM datasets. Examine cognitive impact of BGR tool on architectural processes. (extracted) |
| Reviewer Notes | Strong Category A paper. SUS usability score 74.37/100 from 12 participants validates practical applicability. Key gap: no real-world BIM validation; entirely synthetic data. (extracted) |
| Quality Assessment | 4/6 · rigor: Medium-High · Code/Data: —Benchmark: —Baseline: ✓CV/Split: ✓Ablation: ✓Reproducible: ✓ (extracted) |
| GNN architectures detected | DGCNN · GraphSAGE / SAGEConv · GCN |
|---|---|
| Frameworks / libraries | Deep Graph Library (DGL) · PyTorch · Grasshopper / Rhino |
| Benchmark datasets referenced | CityGML |
| Hyperparameters (regex-detected) | epochs=100 · batch_size=1 · learning_rate=0.0001, 1- · hidden_units=32 · optimizer=Adam optimizer |
| Dataset stats found | 2136 graphs · 171,232 nodes · 1496 graphs · 638 examples |
| Reported metrics + values | Accuracy: 99.8% · Accuracy: 99.69% |
| Quality (heuristic, 6-flag) | 3/6 · rigor: Medium Code/Data: —Benchmark: ✓Baseline: —CV/Split: ✓Ablation: —Reproducible: ✓ |
| Seed of .bib export → verified neighbours | This paper was queried as a Connected Papers seed (41 entries scanned). In-corpus neighbours verified by the bibliometric ranking: [20] 2025 · Iliadis · Comprehensive framework for dynamic energy assessment of… [42] 2022 · Noueihed · Knowledge-based virtual outdoor weather event simulator u… |
|---|
Part of the GML/GNN in AEC Systematic Review (PRISMA 2020) — 112 papers, 2020–2026