Spatial layout design in architecture through deep neural network algorithms

The architectural layout design is a well-known algorithmic problem in computer-aided architectural design. It is the assignment of discrete space elements to their corresponding locations while attempting to satisfy geometrical and topological goals in their layout. This task requires maintaining consistency to ensure that the requirements are met and is exploratory and iterative in nature. The complexity of this problem has encouraged researchers to explore computational approaches for predicting the space layouts. This activity, takes place during the preliminary design phase and is highly significant, as it impacts later stages of the building lifecycle. Therefore, these methods has the potential to enhance the effectiveness and efficiency of spatial layout in design and suggested to make the work of architects easier and faster. Numerous methods have been proposed to solve the space layout design problem. Each one can be viewed as a rule-based strategy that attempts to simulate space layout design using some high-level rules. However, for producing space layout designs, in addition to the quantitative criteria that may be tested and assessed in a logical process, numerous non-quantitative elements also exist. These qualitative criteria are frequently based on a variety of factors and are challenging to describe; thus, hard-coding them would not be possible or effective. It would be best if the program could learn these rules from existing examples. Some potential solutions can be found in the rapidly expanding field of machine learning, which can serve as a tool for decision-making. Deep learning, a subfield of machine learning, can adaptively achieve goals by learning from data and interpreting experiences. The generative adversarial network (GAN), is a deep learning algorithm that has shown remarkable outcomes in the development of 2D designs. In this paper, GAN is applied to generate automated space layouts with given boundaries. A specialized training dataset, comprising 660 existing apartment layouts from Hamadan, is prepared, with each layout labelled using different colours to represent various spaces for training the model. After the model is trained, the boundary lines of 12 new apartments are tested. The performance of the model is also evaluated using two methods: the pixel accuracy measure as the quantitative method and a qualitative assessment by an expert architect based on the evaluation criteria. The results show that the proposed model successfully generates space layout plans from predefined boundaries. This issue indicates its potential for application in other cases and designs. We propose this model as a tool to facilitate the architectural layout design process, enabling architects to quickly and precisely meet client requests particularly in the projects with complex topological constraints.