Hybrid Supervised and Reinforcement Learning for the Design and Optimization of Nanophotonic Structures

  • 2022-09-08 23:43:40
  • Christopher Yeung, Benjamin Pham, Zihan Zhang, Katherine T. Fountaine, Aaswath P. Raman
  • 0

Abstract

From higher computational efficiency to enabling the discovery of novel andcomplex structures, deep learning has emerged as a powerful framework for thedesign and optimization of nanophotonic circuits and components. However, bothdata-driven and exploration-based machine learning strategies have limitationsin their effectiveness for nanophotonic inverse design. Supervised machinelearning approaches require large quantities of training data to producehigh-performance models and have difficulty generalizing beyond training datagiven the complexity of the design space. Unsupervised and reinforcementlearning-based approaches on the other hand can have very lengthy training oroptimization times associated with them. Here we demonstrate a hybridsupervised learning and reinforcement learning approach to the inverse designof nanophotonic structures and show this approach can reduce training datadependence, improve the generalizability of model predictions, and shortenexploratory training times by orders of magnitude. The presented strategy thusaddresses a number of contemporary deep learning-based challenges, whileopening the door for new design methodologies that leverage multiple classes ofmachine learning algorithms to produce more effective and practical solutionsfor photonic design.

 

Quick Read (beta)

loading the full paper ...