Inverse design of 3D cellular materials with physics-guided machine learning

Abstract

This paper investigates the feasibility of data-driven methods in automating the engineering design process, specifically studying inverse design of cellular mechanical metamaterials. Traditional methods of designing cellular materials typically rely on trial and error or iterative optimization, which often leads to limited productivity and high computational costs. While data-driven approaches have been explored for the inverse design of cellular materials, many of these methods lack robustness and fail to consider the manufacturability of the generated structures. This study aims to develop an efficient inverse design methodology that accurately generates mechanical metamaterial while ensuring the manufacturability of predicted structures. To achieve this, we have created a comprehensive dataset that spans a broad range of mechanical properties by applying rotations to cubic structures synthesized from the nine cubic symmetries of cubic materials. We then employ a physics-guided neural network (PGNN) consisting of dual neural networks: a generator network, which serves as the inverse design tool, and a forward network, which acts as a physics-guided simulator. The goal is to generate desired anisotropic stiffness components with unit-cell design parameters. The results of our inverse model are examined with three distinct datasets and demonstrate high computational efficiency and prediction accuracy compared to conventional methods.