Robust inverse-design of scattering spectrum in core-shell structure using modified denoising autoencoder neural...

Area of Science:

• Nanophotonics
• Computational electromagnetics
• Machine learning

Background:

• Neural network-based inverse design offers computational efficiency for nanophotonic devices.
• Conventional methods struggle with robustness and stability against variations in target electromagnetic responses.
• Ensuring approximation capabilities for non-existent exact target responses is crucial.

Purpose of the Study:

• To develop a robust and stable neural network for nanophotonic device inverse design.
• To improve the accuracy of inverse design by addressing input variations.
• To create a network capable of approximating desired electromagnetic responses.

Main Methods:

• A modified denoising autoencoder network was developed.
• The network incorporates a pre-trained model replacing traditional numerical simulations.
• Training involved adding random disturbances to the dataset generated by the pre-trained network.

Main Results:

• The modified denoising autoencoder network demonstrated superior robustness and accuracy compared to fully connected networks.
• The proposed network effectively handles variations in the input target electromagnetic response.
• Successful application in achieving desired scattering spectra for layered spherical scatterers was shown.

Conclusions:

• The modified denoising autoencoder network provides a robust and accurate solution for neural network-based nanophotonic inverse design.
• This approach enhances stability against input electromagnetic response variations.
• The method is flexible and effective for designing nanophotonic devices with specific spectral properties.