Advancing graph-supported machine learning in generative design for architectural engineering

Graph-Supported Machine Learning (GML), including Graph Neural Networks (GNNs) and Probabilistic Graphical Models (PGMs), shows promise for tackling Generative Design (GD) challenges in architectural engineering. However, a systematic review of its use, limitations, and future potential is still lacking. This paper addresses that gap by analyzing 70 peer-reviewed papers, mapping their applications, data sources, and model types. A two-tier analysis identifies key limitations, including small datasets, narrow generalization, and limited integration of physical laws or expert feedback. To overcome these challenges, five strategic directions are proposed: co-evolving data and algorithms, hybrid modeling with Bayesian Networks and GNNs, graph sparsification, human-in-the-loop design refinement, and physics-informed learning. These directions guide the development of more versatile and practical GML models, able to adapt across scales, reduce computational cost, and align with design intent and engineering principles. The findings are intended to foster innovative design practices and advance automation in construction through enhanced computer-aided design.