Synthesizing 3D Morphology from a Collection of Urban Design Concepts

Advisor: Prof. Takehiko Nagakura, Prof. Terry Knight
Reader: Asst. Prof. Justin Solomon, EECS

An urban design proposal guides professionals (architects, developers, and contractors) to build lively urban environments. The core task of urban design involves designing specific 3D urban morphology, aka the collection of building typologies (parcel level), typically executed by a small group of urban designers. A qualified 3D urban morphology requires designers to continuously iterate their individual ideas with multiple stakeholders. However, due to the high costs of visualizing massive building geometries in the process, the current urban design workflow is not able to realize adequate iterations. To avoid deploying a building typology to every parcel of a project manually, a traditional generative design approach such as ESRI’s CityEngine utilizes rule-based systems with parameters of building and zoning code. However, it only provides limited options of building typologies, constructing a bottleneck (Figure 1) of generating 3D morphology from creative design concepts of building typology.

This thesis proposes a machine learning pipeline to synthesize novel 3D morphology from multiple urban design precedents automatically, solving the bottleneck of rule-based generative design above. By training on 3D GIS datasets, the neural network can extract building typologies from a 2D aerial perspective image of urban design precedents as feature vectors in probabilistic latent space. The network has been improved based on urban design workflow, extracting feature vectors of not only building typologies but also heights, land-use, parcel shape, etc. By concatenating these feature vectors automatically or replacing parts of them with designers’ target features (like assigned land-use), this approach is capable of generating novel 3D morphologies.

The deliverable content focuses on data preprocessing, machine learning training result and feature vector demo this semester (Spring 2019). This thesis will be submitted in Fall 2019 with improved performance, user interface, and evaluation.

Figure 1: the bottleneck to generate an urban proposal automatically (outcome example: Morphosis' design for Unicorn Island )