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

Broadband microresonator frequencycombs are being intensely pursuedfor deployable technologies like optical atomic clocks. Spectral features,such as the dispersion in their coupling to an access waveguide, arecritical for engineering these devices for application, but optimizationcan be computationally intensive given the number of different parametersinvolved and the broad (octave-spanning) spectral bandwidths. Machinelearning algorithms can help address this challenge by providing estimatesfor the coupling response at wavelengths that are not used in thetraining data. In this work, we examine the accuracy of three neuralnetwork architectures: fully connected neural networks, recurrentneural networks, and attention-based neural networks. Our resultsshow that when trained with data sets that are prepared by includingupper and lower limits of each design feature, attention mechanismscan predict the coupling rate with over 90% accuracy for spectralranges 6x wider than the spectral ranges used in training data.Consequently, numerical optimization for the design of ring resonatorscan 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 featuresand performance metrics, even greater acceleration may be achievable.

Automated summary

In this study, the accuracy of three different neural network architectures in optimizing the design of ring resonators was examined. The use of attention-based neural networks allowed for predicting the coupling rate with over 90% accuracy for spectral ranges 6x wider than the training data. This significantly reduces computational burden and potential compute time by 6-fold for the design of broadband microresonators.