FB-GAT : A graph neural networks (GNNs) approach to assessing facades’ buildability

Façades always determine the success or failure of a building. A buildable façade design can not only reduce the construction time, cost, and quality defects, but also remove the safety risks, save carbon footprints, and instigate other social and environmental benefits. Traditional buildability assessment methods rely heavily on manual work without tapping into emerging artificial intelligence instruments. This study introduces FB-GAT, a Graph Neural Networks-enabled approach to assessing the buildability of facades by focusing on complex high-rise residential buildings. Our approach starts by collecting representative façade drawings, assigning buildability scores based on expert knowledge and criteria, and representing such drawings as attributed graphs. A graph learning algorithm then encodes the graph data with scores and employs dynamic attention to learn key node-level features. The approach achieved a satisfactory Mean Squared Error of 0.0098 by evaluating on 106 real-life facades of high-rise residential buildings in China. It further identified critical design patterns influencing buildability. This is the first research of its kind to use innovative approaches to rationalize building facades’ buildability in a quantitative and graphical fashion.