Training and evaluating a Co-pilot tool using graph neural networks for generating non-orthogonal building typologies in architectural autocompletion

Abstract

This study introduces a context-sensitive generative artificial intelligence (AI) co-pilot tool designed to assist architectural design by generating predictive suggestions based on Building Information Modeling (BIM) and Industry Foundation Classes (IFC) data. The approach adopted in this study involves the utilization of Graph Neural Networks (GNNs) and Deep Generative Models of Graphs (DGMG) to enhance the synthesis of 3D architectural spatial typologies. To capture spatial and relational patterns between building elements, various GNN architectures where employed, including Graph Convolutional Networks (GCNs), GraphSAGE, and Graph Attention Networks (GATs). The custom training dataset comprised 180,000 subgraphs derived from real-world BIM models, with IFC files converted into heterogeneous graph representations. A combination of multi-label classification and regression techniques was applied to address the complexities of architectural design predictions. The developed co-pilot tool integrates these models within an interactive human-AI workflow, offering users different levels of control. A System Usability Scale (SUS) assessment was conducted for each control mode, demonstrating the potential of the tool to generate context-sensitive design recommendations. Visual evaluation findings highlight the co-pilot’s ability to learn spatial relationships between building elements in 3D and predict contextually appropriate architectural components.