Building morphology generative design : a knowledge-driven paradigm considering comfort, context, and cost

Abstract

Artificial intelligence (AI) is reshaping the creative professions in an unprecedented way. Machines equipped with AI can not only perform repetitive tasks swiftly and accurately but also learn the tacit knowledge hidden inside data and make sensible decisions. Architecture is no exception. As an old profession that is dedicated to creating shelters for human beings across millennia, experience has been passed through generations of designers. The experience can be found both explicitly in technical treatises such as Yingzao Fashi and De Architectura as well as implicitly in building morphologies, structures, and materials. So, to enable machines to understand knowledge has become a research frontier.

Generative design refers to a to-and-fro process in which designers deploy algorithms to produce and analyze design possibilities given user inputs. A twin of generative design paradigms arises. The rule-based paradigm uses relationships among design elements to generate variants, whereas the knowledge-driven paradigm applies machine learning to decode knowledge from past works into models and use them for design generation, evaluation, and optimization. While the former is popular, its reliance on one or several designers may lead to limited possibilities. In contrast, the knowledge-driven paradigm broadens the design frontiers by integrating the knowledge of past works. Nonetheless, the relevant literature is few.

The thesis aims to advance the generative design field built upon the knowledge-driven paradigm. It does so by applying the paradigm in building morphology generative design and considering performances in terms of comfort, context, and cost. Building morphology allows designers to focus on forms and functions over the external ornaments. A three-step workflow: generation, evaluation, and optimization, is implemented. The first trains a machine learning model that can map input design variables to morphological outcomes. The ground-truth data is collected from Hong Kong, with the multivariate Random Forest (mRF) model selected as the learner. Then, thermal comfort, construction cost, and style similarity are evaluated by simulation tools, with univariate Random Forest (uRF) models trained to surrogate. Lastly, the performances are enhanced via a genetic optimizer towards Pareto fronts.

The models present great generative abilities in real-life scenarios. They can derive heterogeneous morphologies based on various design requirements and shapes of sites. The optimization results also demonstrated that the metaheuristic algorithms located Pareto fronts in which there are clear trade-offs among the three objectives. As a result, less than 10% of occupants will be dissatisfied, with 12% construction cost on average being decreased, and less than 18% style similarity being achieved.

The research goes further to discuss how to better incorporate knowledge-driven paradigm in actual design projects. An objective-oriented system architecture is proposed to accommodate models and meet demands from real-life practices. This research points out that the AI can fit into the common design practices. The research also provides empirical evidence for trading off design options that were traditionally considered as hard to assess quantitatively. Finally, this research argues that human designers should less give direct instructions to AI but focus on (1) giving clear performance targets, and (2) updating the models.