Data-driven design of thin-film optical systems using deep active learning.

Area of Science:

• Optics and Photonics
• Materials Science
• Artificial Intelligence

Background:

• Designing optical systems, particularly anti-reflective coatings, is crucial for reducing light reflection.
• Traditional optimization methods for inverse design problems can be computationally intensive and time-consuming.

Purpose of the Study:

• To develop a deep learning-aided optimization algorithm for designing flat thin-film multilayer optical systems.
• To demonstrate the algorithm's effectiveness in creating high-performance anti-reflective gratings and dielectric stacks.

Main Methods:

• Utilized a deep generative neural network, specifically a variational autoencoder, for optimizing photonic devices.
• Applied deep active learning to develop a regression model for predicting optical system reflectivity.

Main Results:

• Successfully designed and constructed a dielectric anti-reflective stack on silicon with an average reflection of 1.52%.
• Achieved lower average reflection compared to recently published experimental results in optical engineering and physics.
• Demonstrated significantly lower computational cost compared to traditional nonlinear optimization algorithms.

Conclusions:

• Deep learning advancements can significantly enhance inverse design capabilities in photonics.
• The developed algorithm offers a computationally efficient and effective approach for designing advanced optical systems.
• The predictive regression model serves as a valuable tool for rapid evaluation and design of optical systems.