Neural Networks for Structural Optimisation of Mechanical Metamaterials

Abstract

Mechanical metamaterials are man‐made designer materials with unusual properties, which are derived from the micro‐structure rather than the base material. Thus, metamaterials are suitable for tailoring and structural optimisation to enhance certain properties. A widely known example for this class of materials are auxetics with a negative Poisson's ratio. In this work an auxetic unit cell is modified with an additional half strut.During the deformation this half strut will get into contact with the unit cell and provide additional stability. This leads to a higher plateau stress and consequently to a higher energy absorption capacity. To achieve the maximum energy absorption capacity, a structural optimisation is carried out. But an optimisation exclusively based on finite element simulations is computationally costly and takes a lot of time. Therefore, in this contribution neural networks are used as a tool to speed up the optimisation. Neural networks are one of many machine learning methods and are able to approximate any arbitrary function on a highly abstract level. So the stress‐strain behaviour and its dependency from the geometry parameters of a type of microstructure can be learned by the neural network with only a few finite element simulations of varying geometry parameters. The modified auxetic structure is optimised with respect to the mass specific energy absorption capacity. As a result a qualitative trend for the optimal geometry parameters is obtained. However, the Poisson's ratio for this optimisation is close to zero.