From Prediction to Design: A Machine Learning Tool for Porous Mechanical Metamaterials

Abstract

Due to their interesting internal architectures, spherical-based porous metamaterials exhibit outstanding mechanical characteristics, naturally minimising stress concentrations. Yet, despite these promising characteristics, they remain underexplored, and the relation between their structural and functional properties is still not well understood. To fill this gap, we exploit machine learning (ML) algorithms to relate the mechanical properties of spherical-based porous metamaterials to their architecture in a two-way fashion. First, we develop artificial neural networks (NNs) to link the stiffness and/or porosity of metamaterials to their structural properties. Then, an inverse design framework is realised by coupling the developed NNs to genetic algorithms (GAs). This strategy allows to efficiently explore the design space and optimise the metamaterial architecture based on specifications on target properties. The proposed tools enable the generation of spherical-based porous metamaterial designs with normalised stiffness [[EQUATION]] between 0.060 and 0.226 and porosity [[EQUATION]] between 0.55 and 0.80. Furthermore, the proposed method identifies geometrical features critical to obtain the desired target properties, specifically the strut thickness [[EQUATION]] and the vertical interconnection parameter [[EQUATION]]. The presented frameworks are intended to be used either as a standalone or coupled with other computational tools to evaluate the utilisation of the developed metamaterials in various engineering scenarios.

Keywords: Spherical-based Porous Metamaterials, Mechanical Metamaterials, Artificial Neural Network, Genetic Algorithm, Inverse Design, Porous materials