Deep Learning-Assisted Design for High-Q-Value Dielectric Metasurface Structures.

Area of Science:

• Optics and Photonics
• Materials Science
• Computational Physics

Background:

• Optical sensing technologies are vital for detecting material property changes in biology, medicine, and food safety.
• Dielectric metasurfaces offer enhanced sensitivity and speed for material detection due to their resonant properties.
• Traditional electromagnetic simulations for metasurface design are computationally intensive and time-consuming.

Purpose of the Study:

• To propose a novel forward prediction network for the amplitude spectrum of dielectric metasurfaces.
• To address the slow computational speed of traditional methods in metasurface design.
• To enhance the efficiency of designing high-quality-factor (Q-value) resonant dielectric metasurfaces.

Main Methods:

• Development of a forward prediction neural network for dielectric metasurface amplitude spectra.
• Utilizing electromagnetic simulations to determine resonant wavelengths.
• Employing transfer learning to adapt the network for near-infrared transmission spectra prediction.

Main Results:

• The proposed neural network achieved a mean square error consistently below 10-3.
• The network completed predictions in under 1 second, demonstrating high-precision capability.
• Transfer learning successfully applied the network to predict near-infrared spectra of high-Q-value metasurfaces.

Conclusions:

• The developed neural network significantly enhances the efficiency of dielectric metasurface design.
• The network demonstrates high precision and speed for spectral response prediction.
• The proposed model can serve as a universal backbone for predicting spectral responses of various dielectric metasurfaces.