Prompts to layouts: Hybrid graph neural network and agent-based model for generative architectural design

Architects need efficient generative methods for handling complex architectural layout design tasks to spare more attention to the aesthetics of buildings. High expertise requirements of the input conditions and the large size of datasets bring challenges for architects using generative architectural design methods. This paper presents a hybrid model that integrates Graph Neural Networks (GNNs) for generating architectural layouts with rational topological relationships based on simple prompts and Agent-Based Modeling (ABM) for reducing the dataset size of model training. The generated layouts achieve a Structural Similarity (SSIM) of 0.82 with a Graph Edit Distance (GED) of 1.67 after training on 150 building samples through several testing scenarios. The hybrid model generates layouts efficiently and avoids impediments to the model generalizability due to excessive usage costs for architects. This paper illuminates how leveraging intelligent algorithms, when enriched with data-driven insights, can bridge gaps in collaboration between architects and generative methods.