Nanophotonic inverse design with deep neural networks based on knowledge transfer using imbalanced datasets

Area of Science:

• Nanophotonics
• Computational Materials Science
• Machine Learning

Background:

• Deep neural networks (DNNs) are increasingly used for nanophotonic inverse design.
• Training DNNs for material selection requires large datasets, posing challenges for design efficiency.
• Existing methods face a trade-off between performance and design time due to data demands.

Purpose of the Study:

• To develop an efficient DNN training method for nanophotonic inverse design using imbalanced datasets.
• To overcome the data-intensive limitations of traditional DNN training in material selection.
• To enable the inverse design of nanophotonic devices with discrete material choices and continuous structural parameters.

Main Methods:

• Proposed a multi-scenario training approach for DNN models with imbalanced datasets.
• Utilized datasets approximately four times smaller than conventional methods.
• Employed a hybrid optimization algorithm combining genetic algorithms and gradient descent for design.

Main Results:

• Achieved a high-precision predictive DNN model with significantly reduced training data.
• Successfully designed multilayer nanoparticles and nanofilms.
• Demonstrated the ability to freely select discrete materials from a library.

Conclusions:

• The multi-scenario training method effectively addresses data limitations in DNN-based nanophotonic inverse design.
• This approach enhances design efficiency and allows for simultaneous optimization of material type and structure.
• The method shows promise for increasingly complex material libraries and nanophotonic device designs.