Generating space layout heat maps with cGAN algorithms in artificial intelligence

Space layout designs have always played a crucial role in the early stages of architectural design. This layout representation implies its topological and geometrical constraints which are the resultant of some subjective and objective agents. These subjective and objective agents are hidden semantics that is embedded in the layouts. In the last few decades, the scholars have proposed several algorithmic methods to solve the space layout design problem. However, defining the function of the layout design have always been the challenging part. In this paper, artificial intelligence data-driven methods are applied to generate synthetic layout design. Instead of using a rule-based optimization model a data-driven prediction modelling approach is applied. Specifically, a conditional generative adversarial network is trained with the prepared dataset. Since the innovation of generative adversarial networks(GAN) with Ian Goodfellow’s influential paper in 2014, different branches of GANs for solving different problems have emerged. Conditional generative adversarial networks (cGAN) is one of the main streams. In this research, cGAN is trained with a specific dataset which could be used for predicting the probability of space allocation in a given boundary. This generated layout could be helpful in the early stage of an architectural design. For this task, a specific training dataset is generated which is used to train a cGAN model. The training dataset consists of 300 existing apartment layouts which are coloured in four different sets of low feature representations. The cGAN model is trained with each of these datasets and the four trained models are evaluated based on the quality of their generated layouts regarding the five pre-defined benchmarks.