|
|
 |
| |
|
|
|
Deep Generative Model for Inverse Design of Terahertz Metamaterial Absorbers Using Conditional Variational Autoencoders |
|
|
|
PP: 873-882 |
|
|
doi:10.18576/amis/200403
|
|
|
|
Author(s) |
|
|
|
Suhaila Abuowaida,
Hamza Abu Owida,
Hamza A. Mashagba,
Suleiman Ibrahim Mohammad,
Azlan B. Abd Aziz,
Radwan M. Batyha,
Asokan Vasudevan,
Mardeni Bin Roslee,
|
|
|
|
Abstract |
|
|
| Metamaterial absorber design has historically been an example of a traditional ``forward approach in which electromagnetic simulation is used to predict absorption based on geometric configuration. The computational expense associated with this iterative optimization process severely restricts the ability to rapidly prototype and explore different regions of the design space. We introduce a novel, physics-based deep generative framework using Conditional Variational Autoencoders (CVAEs) for inverse design of Terahertz metamaterial absorbers that directly generates the required geometric configurations based upon target absorption specifications. Our model develops a probabilistic mapping between desired absorption coefficients and the corresponding optimal combinations of patch width, substrate thickness, and operation frequency. Using training datasets consisting of 9,018 full-wave simulated samples across a range of values for patch widths (20 μm–80 μm), substrate thicknesses (5 μm–20 μm), and frequencies (5 THz–11 THz), our trained CVAE achieved a Mean Absolute Error (MAE) of 2.34 μm when predicting geometric configurations and a reconstruction fidelity (R2) of 0.9876. Moreover, our generative model allows for generation of multiple different design solutions for identical absorption objectives, thereby facilitating multi-objective optimization and satisfaction of fabrication constraints. Experimental results demonstrate that 87% of generated geometries result in absorbed radiation consistent with the specified target absorption within ±3% tolerance, corresponding to a 142× speedup over gradient-free optimization. As such, the proposed framework provides a new pathway towards rapid development of metamaterial prototypes, automation of design workflow processes, and physics-informed generative design processes applicable to electromagnetic systems. |
|
|
|
|
|
|
|
