Deep Learning for Automated Product Design

Abstract Product development is a highly complex process that has to be individually adapted depending on the companies involved, the product to be developed and the related designers. Within this process, the approach and the know-how of the designer are very individual and can often only be described with high effort in a rule-based manner. Nevertheless, numerous routine tasks can be identified that offer enormous automation potential. Machine Learning, especially Deep Learning, has proven an immense capability to identify patterns and extract knowledge out of complex data sets. Autoencoder networks are suitable for the conversion of different 3D input data, e.g. Point Clouds, into compact latent representations and vice versa. Point Clouds are a universal representation of 3D objects and can be derived from various 3D data formats. The goal of the approach presented is to use Deep Learning algorithms to identify design patterns specific to a product family out of their underlying latent representation and use the extracted knowledge to automatically generate new latent object representations fulfilling distinct product feature specifications. A deep Autoencoder network with state-of-the-art reconstruction quality is used to encode Point Clouds into latent representations. In this approach, a conditional Generative Adversarial Network operating in latent space for generation of class-, characteristic- and dimension-conditioned objects is introduced. The model is quantitatively evaluated by a comparison of given specifications and the implemented features of generated objects. The presented findings can be used to support designers in the creation process by automatically proposing appropriate objects as well as in the adaption of future product variants to different requirements. This relieves the designer of time-consuming routine tasks and reduces the effort of knowledge-transfer between designers significantly.