Predicting Broadband Resonator-Waveguide Coupling for Microresonator Frequency Combs through Fully Connected and Recurrent Neural Networks and Attention Mechanism

Broadband microresonator frequency combs are being intensely pursued for deployable technologies like optical atomic clocks. Spectral features, such as the dispersion in their coupling to an access waveguide, are critical for engineering these devices for application, but optimization can be computationally intensive given the number of different parameters involved and the broad (octave-spanning) spectral bandwidths. Machine learning algorithms can help address this challenge by providing estimates for the coupling response at wavelengths that are not used in the training data. In this work, we examine the accuracy of three neural network architectures: fully connected neural networks, recurrent neural networks, and attention-based neural networks. Our results show that when trained with data sets that are prepared by including upper and lower limits of each design feature, attention mechanisms can predict the coupling rate with over 90% accuracy for spectral ranges 6× wider than the spectral ranges used in training data. Consequently, numerical optimization for the design of ring resonators can be carried out with a significantly reduced computational burden, potentially resulting in a 6-fold reduction in compute time. Furthermore, for devices with particularly strong correlations between design features and performance metrics, even greater acceleration may be achievable.