Generative Design (GD) is a design approach that uses algorithms to generate designs. This paper investigates the role of optimisation algorithms in GD process. We study how Pareto Fronts – a classical optimization algorithm output – help designers to browse the variety associated with a design problem. Thanks to the “splitting condition” from design theory, we show that valuable Pareto Fronts for designers are those that allow the exploration of a variety of design parameters without modifying substantially the performance of the designed solution. We call “Splitting Pareto Front” the Pareto Fronts that display this property and investigate how to generate them. We compare, on an electrical battery design problem, two optimization algorithms – NSGA-II and MAP-Elites – based on the design parameters variety they generate. Our results show that MAP-Elites generates Pareto Fronts that are more splitting than those generated by NSGA-II. We then discuss this result in term of the design process: which algorithm is best suited for which design task? We conclude with the importance for future research on Generative Design Algorithms (GDA) to study jointly the functioning of GDA and their expected contribution to the design process.

Generative design (GD) algorithms is a fast growing field. From the point of view of Design Science, this fast growth leads to wonder what exactly is 'generated' by GD algorithms and how? In the last decades, advances in design theory enabled to establish conditions and operators that characterize design generativity. Thus, it is now possible to study GD algorithms with the lenses of Design Science in order to reach a deeper and unified understanding of their generative techniques, their differences and, if possible, find new paths for improving their generativity.

In this paper, first, we rely on C-K ttheory to build a canonical model of GD, based independent of the field of application of the algorithm. This model shows that GD is generative if and only if it builds, not one single artefact, but a “topology of artefacts” that allows for design constructability, covering strategies, and functional comparability of designs. Second, we use the canonical model to compare four well documented and most advanced types of GD algorithms. From these cases, it appears that generating a topology enables the analyses of interdependences and the design of resilience.