Machine Learning Assisted Mechanical Metamaterial Design for Additive Manufacturing

Metamaterials, widely studied for its counterintuitive property such as negative Poisson’s ratio, negative refraction, negative thermal expansion, and employed in various fields, are recognised to provide foundation for superior multiscale structural designs. However, current mechanical metamaterial design methods usually relay on performing sizing optimisations on predefined topology or implementing time-consuming inverse homogenisation methods. Machine Learning (ML), as a powerful self-learning tool, is considered to have the potential of discovering metamaterial topology and extending its property bounds. This work considers the use of Neural Networks (NNs), (De-Convolutional Neural Networks) DCNNs and Generative Adversarial Networks (GANs) to speed up the generation of new topologies for metamaterials. NNs and DCNNs are trained to inversely generate metamaterial designs based on the input target effective macroscale properties, whilst the generator in GANs is expected to output diverse metamaterial microstructures with random noise inputs. This work highlights the potential of data-driven approaches in Design for Additive Manufacturing (DfAM) as an alternative to the time-consuming, conventional methods.