Session: 16-01-01: Government Agency Student Poster Competition

Paper Number: 149850

149850 - Multi-Objective Inverse Design of Impact Resistant Metamaterials Under Varying Strain Rates 

This work pertains to the multi-objective inverse design of impact-resistant metamaterials under varying strain rates. Impact-resistant materials are desirable in a wide range of applications, such as sports, automobiles, military, and aircraft, to name a few. Existing literature deals with refining these structures by performing quasi-static finite element (FE) simulations and then verifying them experimentally, which is a time-consuming and expensive process. Moreover, beyond the low-velocity regime, quasi-static simulations are not representative of real-world dynamic conditions and do not guarantee a similar level of performance. To face these challenges, we propose a multi-objective inverse design approach by combining dynamic impact simulations and machine learning (ML) to overcome the typical trade-off between absorbed energy and transmitted force.

To quantify the impact resistance of metamaterials, we define the peak stress (PS) as the maximum transmitted force and the strain energy density (SED) as the area underneath the stress-strain curve. To enable the discovery of structures that push the boundaries of the PS-SED trade-off, we need a design space encompassing a rich suite of structures, capturing various responses under dynamic loading conditions. Expanding upon the traditional hexagonal honeycomb lattice by encompassing graded lattices, varying in cell wall angle, cell wall thickness, and cell type, alongside the velocity of the impacting projectile, we generate a vast design space comprising approximately 10⁸ possible lattice combinations. To model the complex, high-dimensional relationships between the input lattice parameters and the corresponding PS and SED, deep neural networks (DNNs) are employed. The inverse-design framework consists of two DNNs: the forward model and the inverse model. First, the forward model is trained to learn the mapping between structures and properties. The inverse model is then connected in tandem with the forward model, taking the target properties as input and outputting a vector of lattice parameters. This tandem network is then used to find optimized structures by inputting improved target properties to the inverse model, which then gives us optimized structures corresponding to those properties. This multi-objective optimization approach aims to identify several sets of graded lattices that simultaneously maximize SED and minimize PS, accomplishing the design goal in a fraction of a second without needing costly, conventional optimization methods.

We collect approximately 1000 data points for four different strain rates and train the forward model accordingly. The forward model achieved an R² score of approximately 0.99 and an RMSE of 0.04 on the test dataset, indicating that our dataset is sufficiently robust to capture the complex mapping from lattice parameters to target properties. The inverse model was trained next, utilizing the previously trained forward model to evaluate its outputs. The inverse model achieved an R² score of about 0.98 with an RMSE of 0.06 on the test dataset.

To demonstrate the capability of our framework across various strain rates, we selected two distinct applications: a car bumper and a lacrosse chest protector. Both the car bumper and the chest protector are subject to constraints on maximum strain, PS, and SED. For the car bumper, we consider two scenarios: a high-speed impact representing a pedestrian collision and a low-speed impact representing a car collision. Similarly, for the chest protector, we simulate the impact of a lacrosse ball at speeds of 30 and 50 mph. Our framework can successfully identify optimized structures that surpass all requirements of each application and perform ~50% better than the baseline uniform hexagonal honeycomb, even at equivalent relative density values. The framework is general enough to handle other impact resistance application cases, such as the bicycle helmet. The host of inverse-designed optimized structures pushes the boundary of the current dataset by introducing structures close to the theoretical limit for each application case, i.e., car bumper and chest protector. This data-driven approach combines ML and FE simulations to help accelerate the development of structural materials while providing important mechanics insight into the deformation of the optimized structures.

Presenting Author: Anish Satpati University of California Berkeley

Presenting Author Biography: Anish Satpati is a Graduate Student Researcher (GSR), about to complete his 1st year of PhD at the Department of Materials Science and Engineering (MSE), UC Berkeley. He did his undergraduate studies at IIT Bombay in India in the Metallurgical Engineering and Materials Science (MEMS) Department, from 2019 to 2023. He is enthusiastic and passionate about discovering high-performance structural materials by utilizing data-driven computational tools to accelerate their discovery. He likes to tackle challenges head-on by breaking down seemingly tough problems into easily digestible, smaller parts and solving them one-by-one. Motivated by the fact that each new day brings the opportunity to do breakthrough research for advancement of the field, his goal is to develop a deep domain expertise so that he can teach and inspire young students to keep alive the hunger for mind-boggling scientific discoveries!

Authors:

Anish Satpati University of California Berkeley
Marco Maurizi University of California Berkeley
Rayne Zheng University of California Berkeley