DzAIℕ: Deep learning based generative design

• Published in: Procedia manufacturing Vol. 44; pp. 591 - 598
• Main Authors: Kallioras, Nikos Ath, Lagaros, Nikos D.
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 2020
• Subjects: deep belief networks; Deep learning; Generative design; Machine learning; Reduced order models; restricted boltzmann machines; Solid isotropic material with penalization; Topology optimization; Deep learning; deep belief networks; Solid isotropic material with penalization; Machine learning; Topology optimization; Generative design; restricted boltzmann machines; Reduced order models
• ISSN: 2351-9789; 2351-9789
• DOI: 10.1016/j.promfg.2020.02.251

Generative Design is the methodology for automatic creation of a large number of designs via an iterative algorithmic framework while respecting user-defined criteria and limitations. It is mainly used as a designer assistive tool for the creation of prototypes in sectors like product manufacturing, automotive and aerospace industry while there are also references of use of Generative Design in architecture [1]. The algorithms used are mainly originating from nature-mimicking procedures [2]. Generative Design differentiates from shape optimization due to the fact that it does not focus on finding the optimal design but on being able to propose a large variety of designs that satisfy all designer-defined constraints in an automated procedure. Thus, it can be said that it focuses on proposing initial designs for inspiring the designer. In this work, a methodology for Generative Design called DzAIℕ is proposed. DzAIℕ is based on an algorithmic architecture that combines topology optimization and deep learning methods. Deep learning and specifically Deep Belief Networks (DBN) [3] are used for diversifying through stochasticity, designs generated by topology optimization method Solid Isotropic Material with Penalization (SIMP) [4]. SIMP is used for guiding the design procedure towards a local optimum. The response time for DzAIℕ to propose a series of shapes is minimal as only a small population of SIMP iterations is needed while the time needed for application of a series of DBNs is insignificant. In the current work, a series of examples of computer-generated designs with the use of DzAIℕ are presented in order to validate the proposed methodology.