Abstract
Architectural Design and Urban Planning processes typically involve defining design parameters and comprehend a multitude of stakeholders. This can take a huge toll on the cost and time taken to materialise a project. Typically for projects of scale such as large infrastructure, airports, harbours, mining exploration and even cities, the delay factor can be significant especially in systems that have democratic deciding powers. Early-stage planning typically having low constraining variables gives the room for creative exploratory and creative processes to be performed in an experimental setup allowing breakthrough ideas and concepts to foster. Technology aided conceptual design workflows and “Design Thinking” (Harvard Business Review 2018) frameworks provide a streamlined process for the designer to position their design and evaluate the feedback loop at various stages. The current zeitgeist in technology and software computing, typically are on the increasing scale on the level of abstraction, and usually present simple, intuitive control methods for the end user. Parameterisation and generalisation of variables, methods and actions generate data and enables a whole new level of possibilities to perform intelligent analyses. In contrast, Architectural Design being a combination of sociology, psychology, anthropology, economics, and a reflection of aesthetic, personal, and professional practices only gradually picking up the trend of adapting to changing paradigms. Currently available design tools though capable of intelligent modelling of design parameters, the design data that powers such techniques is often limited by source (Gero and Sudweeks 1995). Furthermore, Architectural and Urban Introduction Data and ML for Early-stage Architectural Design Synthesis 2021 • Singapore University of Technology and Design Ahmed Meeran 5 data are often curated by non-architects and/or designers; Though well annotated and vast datasets such as CubiCasa 5K (Kalervo, et al. 2019) exist the number of such datasets are very minimal, although they are slowly increasing. Hence, data as a paradigm in Technology driven Architectural workflows continues to remain relatively unexplored thus limiting emergent processes such as generative design, contextual understanding, and algorithmic thinking to be performed. The aim of this thesis is to propose a technology powered workflow that leverages machine generated insights to aid architects and planners in their design workflows at the early stage of the design process. Taking the example of Urban Land use Planning, a process has been illustrated that leverages the power of open access data, open-source software, and Machine Learning to gain a site-level and typological understanding of the subject against the surrounding urban fabric. For demonstration, a dataset of the world’s top 1500 airports’ aerial images was computationally retrieved from European Space Agency’s Sentinel API and the usefulness of such pieces of data were materialised into a useful workflow. Using Computer Vision and feature based learning, specific land use classes were trained to assist in taking higher level data driven planning decisions. The lifecycle and the generic nature of the workflow demonstrated makes it a compelling candidate to be incorporated into fast paced environments where different variables are involved in decision making. Another parallel aim of this thesis is to introduce and explore a paradigm of leveraging computer systems (GANs, in specific) to aid in creative exploration of a human designer in the design process. Multiple case studies involving publicly available as well as custom curated datasets with satellite imagery were demonstrated in this thesis to showcase the potential of unsupervised parameter identification and capture evolutionary solution-seeking systems. The findings from this research can be translated into significant use cases with high confidence into current design workflows as tools and packages that can aid in providing context-aware option recommendations, generate novel typologies and help answer typology-entity-site level questions with very minimal requisite information.