INTERACTIVE AI FOR GENERATIVE HOUSING DESIGN BASED ON GRAPH NEURAL NETWORKS AND DEEP GENERATIVE MODELS - Research paper

Oct 26, 2026

Nowadays, housing design complexity has expanded across cost, space, aesthetics, sustainability, circularity, and modularity, stressing traditional labor-intensive workflows. While automatic tools use deep learning for generation, knowledge transfer from human experience remains limited, and data is often not easily understandable or interoperable. Human-in-the-loop interaction is also insufficient for user-centered customization. This research parametrically represents housing designs with graphs and uses graph-representative deep generative models to generate feasible alternatives step by step, supporting interactive human-centered design.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages9

Reading Time23 min

AccessFree to download

Abstract

Introduction

Motivation and challenges in housing design

Prior work in generative floorplan approaches

Graph-based representations and limitations

Proposed direction with heterogeneous graphs

INTERACTIVE AI FOR GENERATIVE HOUSING DESIGN BASED ON GRAPH NEURAL NETWORKS AND DEEP GENERATIVE MODELS Tian Xia 1, Alex Ledbetter 1, Alexandru Bobe 1, Jeroen Hofland 1, Berend Krouwels 1,,Tong Wang 1, Luciano Cavalcante Siebert 1, Paul Chan1, and Jian Yang2 1Delft University of Technology, Delft, Netherlands 2Shanghai Jiao Tong University, Shanghai, China Abstract Nowadays, the complexity of housing design has grown substantially to meet multitudinous requirements (e.g., cost, space, esthetics, sustainability, circularity, and modularity) of diverse stakeholders, which poses great challenges to traditional labor-intensive design processes. Automatic design tools are being developed to assist designers handle tedious work at scale. However, knowledge gaps still exist in harnessing deep learning models to learn from human experience for more efficient design generation while keeping the data understandable and interoperable. Moreover, human-in-the-loop approach is largely neglected in the automatic design tools, which are essential for more customized and usercentered design. This research utilizes graph data to parametrically represent housing designs and graphrepresentative deep generative models for design generation, which provides an interactive design approach for the users at every step. This method enables deep learning models to semantically understand hidden patterns and knowledge in housing designs and facilitate the human-centered design process by returning feasible and parametric housing design alternatives. All codes can be found at: [https://github.com/jlhofland/housing-design](https://github.com/jlhofland/housing-design). Introduction The design process has been identified as one of the most significant elements influencing housing quality (Hamzahet al., 2011) . Traditionally, the creation of housing designs has involved labor-intensive, manual processes, relying on the expertise of designers. This conventional approach has been characterized by iterative design revisions and historical references. However, it faces challenges related to scalability and adaptability to evolving societal needs (Williamson and Wong, 2022) . In response to these challenges, recent advancements in technology and more specifically in the area of Generative AI have introduced the possibility of partially automating the housing design process (Ko et al., 2023) . These advancements have the potential to accelerate design iterations and augment the creativity of the proposed designs (As and Basu, 2021) . While various techniques for automatic housing design generation offer distinct advantages, they also come with some limitations. Generative Adversarial Networks (GANs) have the ability to produce realistic floorplan images (He et al., 2022; Wu et al., 2019) . However, the image-based approaches lack interoperability with common architectural applications (e.g., CAD and BIM)(Ko et al., 2023) . Moreover, it is hard to incorporate various constraints and user requirements into the image- based floorplan generation models. Graph-based methods like Graph Neural Networks (GNNs) can be a complementary approach which has good interoperability due to its parametric nature and compatibility with stateof-the-art generative models (Guo and Zhao, 2023) . Therefore, it is worth noting that a promising approach lies in combining GANs with graph-based methods for their generative ability and interoperability. Numerous studies have incorporated graphs as semantic representation to guide the floorplan generation with GAN, which could be referred to as graph-constrained Relational Generative Adversarial Networks (RaGAN)(Nauata et al., 2021, 2020; Shabani et al., 2022) . By using graphs, the users can more easily input and modify the layout at the semantic level, which provides a user-in-theloop approach and ensures that the model not only produces functional designs but also aligns closely with the user requirements. Bubble diagrams ha

FAQs

What problem does the research address in housing design generation?

It tackles the growing complexity of housing design requirements and the limitations of traditional and automatic tools in capturing human knowledge, supporting interoperability, and enabling interactive human-in-the-loop customization.

How does the proposed method represent housing designs?

It parametrically represents housing designs using graph data, enabling structured encoding of design elements and constraints for generation.

What is the interactive human-centered contribution of the approach?

The method provides an interactive design approach where users receive feasible, parametric housing design alternatives at every step, aligning generated results with user-centered requirements.

Oct 26, 2026

Upload by Dewi_Sukmawati