Hybrid supervised and reinforcement learning for the design and optimization of nanophotonic structures

• DOI: 10.1364/oe.512159
• Publication Date: 2024-02-15T17:00:31Z
• Authors: Christopher Yeung; Benjamin Pham; Zihan Zhang; Katherine T. Foun...; Aaswath P. Raman
• Keywords: Leverage (statistics); Nanophotonics; Testbed

From higher computational efficiency to enabling the discovery of novel and complex structures, deep learning has emerged as a powerful framework for design optimization nanophotonic circuits components. However, both data-driven exploration-based machine strategies have limitations in their effectiveness inverse design. Supervised approaches require large quantities training data produce high-performance models difficulty generalizing beyond given complexity space. Unsupervised reinforcement learning-based on other hand can very lengthy or times associated with them. Here we demonstrate hybrid supervised approach structures show this reduce dependence, improve generalizability model predictions, significantly shorten exploratory times. The presented strategy thus addresses several contemporary challenges, while opening door new methodologies that leverage multiple classes algorithms more effective practical solutions photonic