Estimating and Exploring the Product Form Design Space Using Deep Generative Models

Product forms in quantitative design methods are typically expressed with a mathematical representation such as vectors, trees, graphs, and grammars. Such formal representations are restrictive in terms of realism or flexibility, and this limits their utility for human designers who typically create product forms in a design space that is restricted by the medium (e.g., free-hand sketching) and by their cognitive skills (e.g., creativity and experience). To increase the value of formal representations to human designers, this paper proposes to represent the design space as designs sampled from a statistical distribution of form and estimate a generative model of this distribution using a large set of images and design attributes of previous designs. This statistical representation approach is both flexible and realistic, and is estimated using a deep (multi-layer) generative model. The value of the representation is demonstrated in a study of two-dimensional automobile body forms. Using 180,000 form data of automobile designs over the past decade, we can morph a vehicle form into different body types and brands, thus offering human designers potential insights on realistic new design possibilities.