Session: 17-01-01: Research Posters

Paper Number: 150465

150465 - Automated De Novo Design of Architectured Materials Based on Explanable Artificial Intellegence 

Stochastic microstructures are random in nature and typically fabricated using methods that do not lend themselves to the direct control of microstructure, while engineered architectured Materials, such as metamaterials with periodic patterns, can achieve superior properties compared with their stochastic counterparts, such as the random microstructures found in natural materials. This study focuses on leveraging knowledge learned from stochastic microstructures to facilitate property-driven generative design of periodic microstructures (e.g. cellular metamaterials).

By reviewing the papers, traditional methods which are used to get the features or inspiration from existing designs tend to rely on the judgement of human beings to generate new microstructures, such as brainstorming-based, ad hoc design inspiration approaches, which are inconvenient and unreliable. Instead of them, we propose an eXplainable Artificial Intelligence (XAI)-based framework with the help of Gradient-weighted Regression Activation Mapping (Grad-RAM) to automatically learn critical features from the exceptional outliers (with respect to properties) in stochastic microstructure samples, enabling the generation of novel periodic microstructure patterns with superior properties. By using a dataset consisting of 20000 stochastic samples, we trained some regression models with high accuracy and applied Grad-RAM method on them, so that we can capture the critical features in the original designs and regenerate the new designs.

To test the efficiency of the proposed framework, in this study, this framework is demonstrated on two benchmark cases: designing 2D cellular metamaterials to maximize stiffness in all directions, and to maximize the Poisson’s ratio in all directions. The effectiveness of this framework is demonstrated through the successful generation of novel designs exhibiting superior properties compared to all known samples within the existing dataset. Simulations on ABAQUS are used to calculate the properties for comparison. Besides, according to the comparisons between topology optimizations with different start points, we can conclude that the generated results are closer to the local optimal points. Therefore, this work highlights the transformative potential of XAI in surpassing the limitations of traditional ad hoc design processes, which typically rely on human expertise and brainstorming to transfer knowledge learned from one category of microstructures to inspire novel designs in another category (e.g., from stochastic to periodic). The proposed design framework enables design automation and reduces the reliance on human judgment and brainstorming. The results validate the hypothesis that by learning advantageous microstructure features from stochastic microstructures using XAI, novel periodic microstructure designs created by arranging those advantageous features in a periodic manner can achieve superior properties compared to stochastic microstructures.

In the future work, there will be more cases to be studied, such as minimizing the thermal compliance or getting the negative Poisson’s ratio. With the help of different XAI methods, we can design more kinds of microstructures with different properties.

Presenting Author: Zhengkun Feng University of Connecticut

Presenting Author Biography: Zhengkun Feng is a first-year Ph.D. student in Mechanical Engineering at the University of Connecticut, under the supervision of Dr. Hongyi Xu. His research focuses on developing novel design methodologies for the microstructures of metamaterials, utilizing deep learning techniques. He has finalized the paper as the first author for this representation, showcasing his innovative work.

Authors:

Zhengkun Feng University of Connecticut
Weijun Lei State University of New York at Stony Brook
Leidong Xu University of Connecticut
Shikui Chen State University of New York at Stony Brook
Hongyi Xu University of Connecticut