Inverse Design of Mechanical Metamaterials Using a Point-Cloud-Based Deep Generative Model (2411.19681v1)

Abstract: Mechanical metamaterials have garnered attention for their engineered mechanical properties, enabling control over specific behaviors. Advances in additive manufacturing have expanded design freedom for complex metamaterials. However, the design process remains challenging due to vast design space and numerous parameters. While AI aids design, current approaches are often restricted to predefined, parameterized structures. This study introduces a parameter-free design strategy for 3D mechanical metamaterials using point-cloud-based deep generative networks. A library of widely known metamaterial structures was constructed to train the machine learning model. The trained latent space forms clusters of unit cell topologies with similar properties, enabling efficient exploration and smooth interpolation. Additionally, mechanical properties can be predicted more faster than with conventional methods. This approach created metamaterials with targeted properties, unrestricted by parameterized constraints. Computational and experimental validations confirmed alignment with desired properties within acceptable error margins. We believe this work significantly enhances design flexibility in AI-driven metamaterials, expanding their potential applications across various fields.