This paper presents a comparison between human-defined and AI-generated design spaces through simple optimisation applications.

A design space is a formal expression of a design idea. It is constructed by selecting a set of variables, which limit the search for suitable solutions to a design problem within a specific range of options. Most computational approaches to structural design are based on parametric modelling, which require the definition of a design space, and therefore an analytical formulation of a design idea. In structural optimisation, such approaches tend to limit the search for optimal solutions to a subset of the entire space of design possibilities, and do not necessarily prompt the designer’s creativity.

Recent AI models, such as Variational Autoencoders (VAEs) (Kingma and Welling, 2014), have the potential to overcome some of the limitations described above. VAEs can construct design spaces by extracting implicit design variables from a dataset of design solutions. Such variables result from a learning process and are conditioned exclusively by the characteristics of the dataset, rather than by a human-formalisation of design thoughts.

A VAE has been trained on an artificial dataset of shell structures to construct a design space, which has then been compared with a design space constructed through the explicit definition of design variables. The comparison has been performed by analysing the diversity of the solutions retrieved from both design spaces in two optimisation applications.

The comparison demonstrates that optimisation based on AI-generated design spaces results in a greater diversity of design outputs than the predictable solutions provided by optimisation based on human-defined design spaces. Furthermore, such design outputs respond better to the selected performance criteria.