Deep Learning-Accelerated Designs and Characterizations of Mechanical and Magneto-Mechanical Metamaterials for Shape Morphing and Tunable Properties

Speaker: Chunping Ma

Abstract:

Metamaterials are deliberately architected artificial materials that can achieve unconventional properties not observed in nature, showing potential for various applications. Mechanical metamaterials are a new branch of metamaterials using geometry designs to control mechanical properties such as stiffness, deformation, and energy absorption.

To date, most of the research on mechanical metamaterials considers an array of unit cells distributed in a uniform pattern, and the properties of those mechanical metamaterials are restricted by the unit cell structure. By introducing multiple unit cells into the array with non-uniform patterns, a much wider variety of mechanical properties become possible. However, such non-uniform mechanical metamaterials with extensive design domains bring challenges to the design, especially when specific target properties are desired. Motivated by the development of deep learning, we develop a framework based on feedforward neural networks (FNN) to systematically explore a large design domain of non-uniform mechanical metamaterials with nonlinearity in material, geometry, and boundary condition, realizing the mechanical response curve predictions of non-uniform patterns and the inverse designs for given target response curves.

But for conventional mechanical metamaterials, their properties are significantly confined by the original geometries and lack in-situ tunability. Thus by a direct ink writing (DIW) technique, we combine hard-magnetic soft materials (MSMs) and hard-magnetic shape memory polymers (M-SMPs), which demonstrate superior shape manipulation performance by realizing reprogrammable, untethered, fast, and reversible shape transformation and shape locking in one material system, to develop magneto-mechanical metamaterials that are capable of shifting between various macroscopic mechanical behaviors such as expansion, contraction, shear, and bending under cooperative thermal and magnetic actuation, enabling wide-range in-situ tunability.

However, considering the massive design domain of the magneto-mechanical metamaterials due to the material and structural programmability, a robust inverse design strategy is desired to create the magneto-mechanical metamaterials with preferred tunable properties. In this research, we develop an inverse design framework where a deep residual network (ResNet) replaces the conventional finite-element analysis (FEA) for acceleration, realizing metamaterials with predetermined macroscopic strains under magnetic actuation. For validation, the DIW for the MSMs is adopted to fabricate the designed complex metamaterials, and the results from the inverse designs, FEA, and experiments are well-matched with each other.

Overall, this research boosts the design ability of both mechanical and magneto-mechanical metamaterials and paves the way for functional programmability at a system level.

Zoom Link: https://osu.zoom.us/j/94883422414?pwd=ZWFSYmxqY0lJZGJZMzZqcnBxS05RUT09.