Inverse Design of Mechanical Metamaterials

This study explores the use of data-driven approaches for automating the engineering design process with a focus on the inverse design of mechanical metamaterials, which are materials with exceptional properties and unique structures that have gained increasing attention in the field of materials science in recent years. Metamaterial conventional design process is slow and requires iterative optimization, hindering productivity and increasing computational costs. However, by using data-driven design methods, significant time and resource savings can be achieved. The aim of this study is to tackle the challenges of developing an efficient inverse design methodology, effectively representing the data, and ensuring the manufacturability of predicted mechanical metamaterials. The study employs a physics-guided neural network (PGNN) to handle complex parametrized datasets and assesses its effectiveness through two case studies of mechanical metamaterial inverse design. In the first case study, the study develops a data-driven approach to synthesize 2D shape-morphing structures using a hexagonal unit-cell with a curved beam as the fundamental component. The data were established by constraining the theoretical expressions that describe the elastic behavior of a single unit-cell to ensure manufacturability. The findings demonstrate that the proposed methodology is capable of synthesizing shape-morphing structures directly from a target shape with remarkable computational efficiency within an expedited timeframe of a few milliseconds. To validate the performance of the synthesized design, we conducted both finite element analysis simulation and tensile test. The validation results were in accordance with the predictions made by the data-driven approach. In the second case study, we investigate the inverse design of cellular truss materials, which poses a challenge in finding a parametrization method that covers a broad range of elastic stiffness behaviors. To address this, we created our own comprehensive dataset that encompasses a wide range of mechanical properties by applying rotations to cubic structures synthesized from the nine cubic symmetries of cubic materials. The study uses a PGNN to match unit-cell parameters with desired anisotropic stiffness components. The results of the inverse model are analyzed using three different datasets and show strong computational power and prediction accuracy. However, limitations are noted, as the training data must encompass a broader range of mechanical properties to be considered as a general approach.