PISC-Net: A Comprehensive Neural Network Framework for Predicting Metasurface Infrared Emission Spectra

Area of Science:

• Electromagnetic Wave Control
• Metamaterials and Nanophotonics
• Artificial Intelligence in Physics

Background:

• Multifunctional metasurfaces offer advanced control over electromagnetic waves.
• Conventional metasurface design relies on computationally expensive numerical simulations.
• Accurate prediction of infrared radiation spectra is crucial for metasurface applications.

Purpose of the Study:

• To develop a rapid and accurate prediction method for infrared radiation spectra of metasurfaces.
• To reduce the computational cost associated with traditional metasurface design.
• To enhance the generalization capacity of prediction networks for metasurface characteristics.

Main Methods:

• Introduction of the Physical Insight Self-Correcting Convolutional Network (PISC-Net).
• Integration of parameter-communication modules and a priori knowledge for physical mechanism recognition.
• Development of a data set construction strategy for precise absorption band tuning (3-14 μm) and use of transfer learning for large-period metasurfaces.

Main Results:

• PISC-Net achieves rapid prediction of infrared radiation spectra with high generalization capacity.
• The network accurately predicts experimental outcomes using only simulation data, validated by comparative analysis.
• The average mean square error for predictions is less than 4%.

Conclusions:

• PISC-Net offers a computationally efficient alternative to traditional simulation methods for metasurface design.
• The proposed approach demonstrates the potential of AI in accurately predicting experimental metasurface behavior.
• This work facilitates cost-effective development and precise tuning of metasurfaces across a wide spectral range.