Decoupled Design of Hybrid Mechanical Metamaterials via Ensembled Deep Learning

Abstract

Regarding the design of mechanical metamaterials, both periodic unit cells and irregular structures with specific continuity have demonstrated promising application potential. More importantly, material distribution-based design methods also provide a representative perspective. However, existing studies rarely associate mature unit cells with irregular structures while simultaneously considering the influence of material distribution. This hybrid design problem warrants further investigation and holds significant potential for expanding the design space of mechanical metamaterials. This study proposes an inverse design strategy capable of accounting for diverse unit cells and multiple materials in a sole metamaterial design with targeted macroscopic mechanical stiffness. An ensembled deep learning model with variational autoencoders and artificial neural networks is constructed to decouple structural and material contributions to overall mechanical properties, which facilitates the independent design of unit cell and material distribution for targeted properties. Integrating the virtual growth algorithm, the proposed method addresses critical challenges in geometric continuity among various types of unit cells. Accurate reconstruction and prediction of hybrid distributions are realized, with SHAP analysis confirming effective decoupling of structural and material influences on the targeted metamaterial design. Final design targets show excellent accuracy of homogenized properties, indicating the efficacy of our approach. The proposed workflow pioneers a novel decoupled approach for designing mechanical metamaterial with hybrid unit cells and multiple materials, setting a foundation for applications in complex mechanical systems and complicated inverse design problems.