Hybrid inverse design scheme for nanophotonic devices based on encoder-aided unsupervised and supervised learning

Area of Science:

• Nanophotonics
• Metamaterials
• Machine Learning

Background:

• Machine learning is used for nanophotonic device inverse design.
• Complex targets, like multi-peak spectra, pose challenges such as overfitting.
• Existing data-driven methods struggle with intricate design requirements.

Purpose of the Study:

• To develop an improved inverse design scheme for nanophotonic devices with complex spectral targets.
• To address overfitting issues in data-driven inverse design approaches.
• To enhance the efficiency and accuracy of designing metamaterials.

Main Methods:

• A hybrid inverse design scheme combining supervised and unsupervised learning was proposed.
• Clustering algorithms and an encoder model were introduced for data preprocessing.
• The scheme was verified using a metamaterial designed for tunable dual plasmon-induced transparency.

Main Results:

• The hybrid scheme significantly reduced the mean squared error (loss function) by over 51% for both training and test datasets.
• Performance was compared against traditional artificial neural networks (ANNs) trained on the entire dataset.
• Effective data preprocessing using clustering and encoding improved model accuracy.

Conclusions:

• The proposed hybrid inverse design scheme offers an efficient improvement for nanophotonic device design tasks.
• The method effectively handles complex spectral targets, overcoming limitations of previous approaches.
• This work paves the way for more sophisticated inverse design in nanophotonics.