Predicting Non-functional Nodes of Floorplan via Graph Neural Network (GNN)

Abstract

This paper presents an innovative approach to automating the floorplan generation process by employing Graph Neural Networks (GNN) to facilitate the transition from functional assignment lists to bubble diagrams and eventually to floorplan graphs. In recent years, there have been many studies on the interconversion of floorplan graph and layout design. However, these studies usually mix up floorplan graph and bubble diagram, despite their distinct roles in representing spatial and functional relationships, respectively. To address this disparity, we introduce a research framework comprising three main steps. First, we generate the CubiBubble5k dataset, which encompasses bubble diagrams and functional lists, drawing on the existing CubiCasa5k and CubiGraph5k datasets. Next, we train a GNN to transform design assignments into structured graph data, utilizing functional lists as input and bubble diagrams as output. Subsequently, we train another GNN that predicts and inserts non-functional spaces, such as corridors and anterooms, into purely functional bubble diagrams, using bubble graphs as input and floorplan graphs as output. We assess the performance of both GNNs and, by integrating our framework with the established graph2plan study, successfully demonstrate the generation of real-world floorplans from project task books. Lastly, we conduct case studies to validate the feasibility of our proposed framework. We use the existing graph2plan platform to visualize the impact of our algorithm on the final layout.