Generating new cellular structures for additive manufacturing through an unconditional 3D latent diffusion model

Abstract

Advances in additive manufacturing (AM) have facilitated the fabrication of cellular structures inspired by those in the natural world. But the design of complex, tessellating cellular structures remains a challenge for human designers, and only a small number of geometries, defined either by connected walls or struts, or by surface equations, have been investigated. This study introduces generative deep learning to the problem, with the aim of synthesising novel cellular geometries producible by AM. Our unconditional 3D latent diffusion model (U3LDM) explores the design space from a new class of training data comprising 10,650 unit cells. A critical task involved developing a varied set of cell geometries based on random permutations of trigonometric surface equations. This was coupled with a stringent set of pass/fail tests to ensure the generated structures possessed structural connectivity and could tessellate in 3D. The new cellular structures were analysed numerically using finite element analysis, fabricated by polymer AM, and subjected to compression tests to verify their manufacturability and mechanical properties. Results indicate that the U3LDM is capable of generating new ‘unseen’ cellular structures with geometries and mechanical properties consistent with those of the training specimens. This method also demonstrates the potential universal technique for creating nature-inspired and AM-manufacturable structures beyond the currently limited set of human-derived geometries.