Machine learning accelerated design of auxetic structures

Abstract

Auxetic metamaterials exhibit unusual expansion perpendicular to the direction of tensile loading, a behavior known as negative Poisson's ratio (NPR). The microstructures of auxetic metamaterials require rational design. It is a significant challenge when optimizing auxetic cellular structures solely through finite element calculations (FECs) due to the vast number of possibilities. In present work, two machine learning (ML), Evolutionary Computation (EC) and Artificial Neural Network (ANN), are employed to accelerate the design and optimization process. During ML training, auxetic cellular structures are encoded using binary sequences consisting only of 0s and 1s. FECs are conducted on some auxetic cellular structures to generate initial data, which are then used to train the EC and ANN models for efficient prediction of the optimal auxetic structure. In addition, the ANN model demonstrates a search speed at least five orders of magnitude faster than FECs when exploring the structure space, outperforming EC in the present task. Finally, the optimal auxetic structures are manufactured by using 3D printing. Their effective Poisson's ratios closely match those obtained from FECs. The present work showcases the powerful capabilities of ML in the design of metamaterials.