Site planning automation using a novel deep learning-based generative model

With the rapid development of artificial intelligence and computing power, deep learning-based generative methods have begun to be applied to site planning automation in recent years. Compared to typical generative methods that overly rely on expert knowledge, these state-of-the-art learning-based methods can automatically capture urban characteristics of real-world design cases and generate site planning solutions compatible to existing urban structures. However, the application of deep generative methods in site planning automation is still in its infancy with unaddressed limitations: (1) Most studies directly use typical deep generative methods for site planning; however, these methods are designed for general generative purpose, making their results less performative; (2) Existing studies usually adopt an image-to-image conversion process, overlooking the impact of various site/design conditions/constraints (e.g., land-use type, building attribute) on site distribution, leading their mod el/results less practical. To address the limitations, this study proposes a novel deep generative model based on generative adversarial network for automated site planning generation. The proposed model is applied to New York City as a case study to synthesize site solutions in census blocks. Both qualitative and quantitative results demonstrate that the proposed model can generate high quality planning solutions satisfying various design requirements. The model is expected to effectively synthesize many site alternatives in real-time for human system interaction and informed decision-making at the early design stage.