Neural networks based inverse design of inhomogeneous tetrachiral honeycombs for desired deformation

Abstract

The challenge of achieving efficient inverse design for honeycomb structures with desired deformations has persisted. To address this, a machine learning framework including two neural networks is introduced, with one used for sensitivity analysis and dataset generation, while the other for inverse design. A tetrachiral honeycomb structure is parametrically modeled using Python scripts and subsequently analyzed with finite element method (FEM) software. A dataset mapping unit cell parameters to honeycomb deformations is fabricated by FEM for training a forward neural network, which has an R-squared value of 0.9680. Based on this trained neural network, four high sensitive parameters were selected for inverse design by sensitive analysis. Then, a dimension-reduced dataset is created to train an inverse neural network with an mean R-squared value of 0.9909. Finally, experimental verifications were performed, which demonstrates an excellent agreement within the design domain. This approach offers promising potential for tailoring honeycomb structures with desired deformation, while also enabling the inverse design of metamaterials with customized properties.