Enhancing Architectural Plan Generation with Machine Learning and Space Syntax Analysis for Optimized Spatial Configuration

Since recent years till now, a lot of research has been conducted on generative design to help architects in the design process to save time and cost, and produce multiple and better solutions for a given problem. Such solutions have been generating architectural plans based on users’ preferences, generating architectural plans that reduce the dependence on energy for cooling by improving natural ventilation, and generating façade systems that reduce harmful sun light and increase natural light. However, there are a few number of research studies that consider the quality of spatial configuration in terms of privacy hierarchy and the spatial relationship of the automatically generated architectural plans. Spatial configuration is one of the most important features in the architectural design process, and using generative techniques in relation to Machine Learning (ML) in this process has become a requirement in recent years. Although a lot of studies have been carried out about shape grammar and ML relationships, there is not a study which combines spatial configuration using Space Syntax (SS) with ML, which can create a potential for this requirement. Therefore, in this paper, a computational framework has been developed to evaluate the spatial configuration of the generated architectural plans by training a supervised neural network on some spatial feature values of three Syrian houses (post-independence from the French colonization period). These values have been gained by analyzing these houses using the DepthmapX software, which is based on Space Syntax theory. The trained model has been tested on another Syrian housing plan from the same typology. The outcomes of the study demonstrate the potential of the trained model to predict the suitable space function with few errors caused by the strong similarity in spatial features of some spaces and the lack of training samples. The trained model can then be integrated into any plan-generating algorithm or used as a separate tool to enable architects to enhance their spatial configuration in the early design stage. Although the trained model is still under development without accomplishment, it creates a base for further investigations in terms of spatial conditions and ML.