Urban design is a recursive, collaborative process shaping cities, streets, and places (Carmona, 2009). This process typically begins first in analysing its wider context (social, economic, environment), second in understanding the site, third in developing designs and lastly in evaluating the design before moving onto architectural design. Therefore, urban design requires a high degree of continuous feedback and design development with different stakeholders and practitioners. A key question that arises is - how can data science and AI methods be used in the urban design process? 

The author argues that data science, rather than replacing designers and planners, can be integrated into every stage of urban design, thereby assisting us through the design process. This idea can be summarized as “machine-in-the-urban-design-loop” or “machine-assisted-urban-design”.

A Dynamic Understanding of Cities through Knowledge Discovery

The first stage of machine-assisted urban design is knowledge discovery, which refers to the process of generating new insights from urban databases.In urban design, modern data science methods enable the translation of tables (e.g. census), text and imagery into new insights. Early examples of these techniques include applying unsupervised machine learning methods for geodemographic classification, which enables learning about the distribution of population across an area. Recent machine learning advances allow richer discovery. For example, large-scale mobility data offers new opportunities to study pedestrian flow and social interactions, as well as enhancing our understanding of segregation and navigation. Computer vision methods can even be applied on remotely-sensed imagery and point cloud data for various purposes. For example,  to quantitatively evaluate street frontages (Law, 2020). An example can be seen below,  where a neural network model has been built to infer automatically the activeness of a street view (Law, 2020).

More Affordable and Accessible Digital reconstruction

The second stage concerns the use of advanced digital reconstruction methods for site analysis and augmented urban design. In the past, digital reconstruction required substantial effort in surveying the site in great detail. With technological advancements, two opportunities have emerged. The first is the increasing availability of low-cost high resolution 2D images and 3D point cloud sensors on consumer and professional grade handheld mobile devices. The second is advancements in neural rendering where highly accurate 3D environments can be re-constructed from a sparse set of 2D images. Neural rendering technologies can be combined with augmented reality (AR) and tools like Unreal Engine 5, to allow designers to immerse themselves in a site and iteratively develop their designs. The training of future urban designers to use these technologies will be an important facet for urban design pedagogy.

Design Development with Generative Methods

The third stage is the application of generative models in urban design development. Traditional methods in urban generative design are creating rule-based systems and use mathematical optimisation. Aside from these, there is a growing interest in using machine learning techniques such as Stable Diffusion or Reinforcement Learning, especially for  design ideation. We envisage that new technologies that combine computer vision (stable diffusion, 3D geometric ML) and large language models (LLMs) will become an intrinsic part of the urban design ideation process. The seamless integration of these advanced tools in human-computer interactive design is beginning to emerge as a research direction where LLM agents can be used to automate design ideation.

Design Evaluation Using Theory-driven Causal Models And Simulations

The fourth and final stages of machine-assisted urban design focuses on the use of data science in design evaluation. While “black-box” models can help predict the social-or-economic utility of a new design, some might argue that these parameters may not be inherently causal. To address this, a more cautious approach involving causal models and simulations is necessary. One example is to link the causal mechanism back to theory. A classic example is space syntax research which has a theoretical foundation that associates its spatial configuration to pedestrian movement, land-use and even land value via a feedback loop. This type of model has been used for evidence-based urban design for many years and has continued to influence the design of many built projects. Empirically spatial network analysis and agent-based models have been constructed based on theories in simulating pedestrian behaviour to assess impacts of infrastructure changes and navigation.

Final Thoughts

This article has explored how data science can support urban design — from knowledge discovery and digital reconstruction to generative design and design evaluation. Rather than replacing designers and planners, data science can be integrated into every stage of urban design, and assist them in various ways. At the same time, to make full use of these tools, significant challenges lie ahead. These include the lack of multi-modal data, limited understanding of causal-and-neuro-mechanisms in urban design and the potential misuse or over-reliance on these methods in pedagogy and practice. Urban design is both an art and a science. As data science tools become more powerful, it is imperative that designers use them responsibly, being grounded in evidence while fostering creativity. With thoughtful integration, these technologies can help create more adaptive, inclusive, and beautiful urban environments.