Stochastic and multi-objective design of photonic devices with machine learning

Area of Science:

• Photonics and optical engineering
• Computational physics
• Materials science

Background:

• Photonic devices require complex designs with many parameters.
• Existing optimization methods often ignore fabrication uncertainties, impacting real-world performance.
• Stochastic quantities critically influence experimental device outcomes.

Purpose of the Study:

• To develop a novel approach for stochastic multi-objective design of photonic devices.
• To integrate unsupervised dimensionality reduction and Gaussian process regression for design optimization.
• To enable comprehensive analysis of device trade-offs considering both deterministic and stochastic factors.

Main Methods:

• Combined unsupervised dimensionality reduction with Gaussian process regression.
• Incorporated stochastic quantities (fabrication uncertainties) into the design process.
• Investigated surface gratings for fiber coupling on a silicon-on-insulator platform.

Main Results:

• Identified promising alternative designs and modeled their statistical responses.
• Analyzed 86 designs, revealing significant differences in yield and worst-case performance due to variability.
• Discovered marked differences in fiber coupling efficiency and back-reflections when considering uncertainties.

Conclusions:

• The proposed approach enables efficient identification of robust photonic device designs.
• Pareto frontiers demonstrate optimized device robustness against fabrication variability.
• Offers a powerful tool for designing photonic devices with stochastic figures of merit.