Modelling the relationships between ground and buildings using 3D architectural topological models utilising graph machine learning

Historically, architects have established different approaches to constructing their buildings on the ground. Classifying the building/ground relationship enables the architect to make informed design decisions during the early design stages. Manual handling of this task is time-consuming, complex as well as prone to errors. This paper leveraged Machine Learning (ML) methods to overcome this difficulty by applying Graph Machine Learning (GML) to 3D topological models, to classify the building and ground relationship. The paper workflow comprised two stages. The first stage involved generating 3D synthetic architectural precedents and created a dataset of their dual graph using Topologic, which is software that computes the spatial relationships between elements. The second stage ran the Deep Graph Convolutional Neural Network (DGCNN) using PyTorch, which is a Python machine learning library developed by Facebook. The paper’s results demonstrate that the system effectively classifies the relationship between building and ground, with the ability to predict a new previously unseen architectural building/ground relationship with high accuracy measurement that aligns with DGCNNs benchmark graphs. The paper concludes by reflecting on the advantages and disadvantages of generating a sizeable synthetic dataset with embedded semantic topological graphs as a formal design method, in addition to outlining future work