Session: DAC-11-01-Design of Engineering Materials and Structures

Paper Number: 89722

89722 - Efficient Design of Acoustic Metamaterials With Design Domains of Variable Size Using Graph Neural Networks 

Most metamaterial systems are designed with periodic unit cells to make the underlying design problem more tractable.  Shifting to nonperiodic unit cells enables a broader range of physical properties at the expense of higher dimensional design spaces of variable size associated with the adjustable quantity and size of physical features.  Representing the physical behavior of these systems with metamodels can enhance the efficiency of the design process, but several challenges must be overcome.  Training metamodels for high-dimensional systems requires large volumes of data, and in nonperiodic systems where the quantity and arrangement of structural features is variable, the metamodels must be valid for systems with a broad range of dimensionalities.  Furthermore, in acoustic and dynamic applications, responses that are sensitive with respect to frequency compound the issue by requiring dense sampling throughout the spectrum. This paper presents a method to address these challenges by representing these systems as graphs and training purpose-built neural network architectures to update the graphs from state to state. Encoding graph states before—and decoding after—calling the state update functions enables the update functions to maintain generality with respect to dimensionality. The trained update functions are then applicable to systems with dimensionalities beyond those frequently observed in training. By skewing training samples toward lower-dimensional systems that are less computationally expensive to simulate, the computational expense of gathering training data can be reduced with minimal loss of accuracy in predicting the dynamic behavior of higher-dimensional systems. The method is demonstrated by designing an asymmetric acoustic absorber.

 

 

Presenting Author: Carolyn Seepersad Univ Of Texas At Austin

Presenting Author Biography: Dr. Carolyn Conner Seepersad is the J. Mike Walker Professor of Mechanical Engineering and the director of the Center for Additive Manufacturing and Design Innovation at The University of Texas at Austin.

Authors:

Tyler Wiest The University of Texas at Austin
Carolyn Seepersad Univ Of Texas At Austin
Michael Haberman The University of Texas at Austin