Multi-Story Floor Plan Generation from Building Volumetric Design Using Graph Neural Network

EasyChair Preprint 15962

Abstract

The automatic design of architectural floor plans using deep learning has been widely studied to assist architectural design. Traditionally, floor plans generated by deep learning have been limited to single floors. Recently, research has been developing the use of graph neural networks (GNNs), which applies deep learning to graph data to generate building volumes that consider the in-building spatial use. Although these studies aim to generate new building volumes, practical architectural design often requires the generation of floor plans within predefined building outlines, constrained by various legal and regulatory requirements. This study proposes a method for generating multi-story floor plans in a given building volume by using a graph convolutional network (GCN), which adapts convolutional operations to graph data representing the given building volume. The implemented GCN model successfully predicted, with an accuracy of 74.66%, the spatial use class for each node within the graph representing the building. This enables the generation of detailed floorplans across multiple floors. This research contributes to the design support of multi-story floor plans in a given building volume. Moreover, when integrated with the latest 3D generative AI technologies, this approach promises to advance the automatic creation of 3D building models with comprehensive interior designs, starting from scratch in volumes initially devoid of any interior information.

Keyphrases: 3D building layout, Floorplan generation, Graph Neural Network, deep learning, generative design

@booklet{EasyChair:15962,
  author    = {Atsuya Ogura and Tomohiro Fukuda and Nobuyoshi Yabuki and Taro Narahara},
  title     = {Multi-Story Floor Plan Generation from Building Volumetric Design Using Graph Neural Network},
  howpublished = {EasyChair Preprint 15962},
  year      = {EasyChair, 2025}}