Publication

Hyperautomation Artificial Intelligence

Deep Learning-based Inverse Design Framework: A Case Study of a Defective Phononic Crystal for Narrowband Filtering

본문

Conference
Asian Congress of Structural and Multidisciplinary Optimization (ACSMO) 2024
Author
Donghyue Lee, Soo-Ho Jo, and Byeng D. Youn
Date
2024-05-21
Presentation Type
Oral
Abstract
 

This study presents a groundbreaking deep learning-based inverse design framework for phononic crystals (PnCs) with defects, utilizing an innovative approach known as surrogate-assisted conditional generative adversarial networks  (SACGANs).  The  proposed  method  significantly  enhances  the  generalization  capabilities  of  traditional conditional generative adversarial networks (CGANs), which are commonly employed in deep learning-based inverse design  processes.  The  enhancement  is  achieved  through  a  strategic  integration  of  advanced  loss  functions,  the implementation of pseudo-labeling techniques, and the adoption of sequential training methodologies, all of which are harmonized with sophisticated surrogate models. This intricate integration culminates in exceptional generalization performance, enabling the production of more feasible and practical designs. This advancement is particularly notable in the complex field of narrowband filter design within the realm of phononic crystals, where precision and specificity are crucial.

Furthermore, this study pioneers the inverse design challenge of PnCs with double defects, an area that had not been previously explored. By venturing into this new territory, the study demonstrates the immense potential of deep learning-based  methods  in  tackling  intricate  and  complex  design  challenges  that  traditional  methods  may  find insurmountable. This exploration into double-defect scenarios opens up new avenues for research and application in phononic crystal design.

The validation results of this study are particularly noteworthy. The SACGAN framework exhibits a marked superiority over other prevalent models such as Conditional Variational Autoencoders (CVAE), traditional CGAN, and CWGAN-GP (Conditional Wasserstein Generative Adversarial Network with Gradient Penalty). This superiority is evident in the framework's ability to generate more accurate, feasible, and innovative designs, especially in scenarios involving double defects.

Moreover, the performance of the SACGAN-introduced deep learning-based inverse design framework is noteworthy, not only for its exceptional capabilities in single-defect scenarios but also for its unparalleled performance in more complex double-defect scenarios. In these scenarios, the SACGAN framework not only matches but often surpasses the capabilities of traditional genetic algorithms, which are renowned for their effectiveness in optimization and design tasks. This achievement marks a significant milestone in the field of phononic crystal design, showcasing the potential of advanced deep learning techniques to revolutionize traditional design and optimization processes.

In conclusion, this study not only provides a novel and effective tool for the design of phononic crystals with single and double defects but also sets a new benchmark in the field of deep learning-based inverse design. The SACGAN framework,  with  its  advanced  capabilities  and  exceptional  performance,  paves  the  way  for  future  research  and development in this exciting and rapidly evolving field.