Machine learning with knowledge constraints for design optimization of microring resonators as a quantum light source

Abstract

Abstract With careful design and integration, microring resonators can serve as a promising foundation for developing compact and scalable sources of non-classical light for quantum information processing. However, the current design flow is hindered by computational challenges and a complex, high-dimensional parameter space with interdependent variables. In this work, we present a knowledge-integrated machine learning framework based on Bayesian Optimization for designing squeezed light sources using microring resonators. Our model, after only 5 optimization rounds, identified two optimal structures with distinct cross-sectional areas and radii (65 $$\:\mu\:m$$ and 110 $$\:\mu\:m$$ ), achieving escape efficiencies over 90% and on-chip squeezing levels of 7.48 dB and 9.86 dB, respectively. Our results demonstrate that by adaptively finding the coupling coefficient through BO, the model has identified optimal points in the over-coupled regions with superior performance. This optimization model is developed specifically for single resonators made of silicon nitride. However, its applicability extends beyond this, and it can be used to model structures with auxiliary rings or other materials like silicon carbide. Our approach is expected to streamline the design of other integrated photonic components, including Mach-Zehnder interferometers and directional couplers, for applications in quantum photonic circuits and optical neural networks.

Keywords

• Quantum light sources
• Integrated photonics
• Microring resonators
• Machine learning
• Bayesian optimization