Generative Neural Networks for Kerr Combs

Light in integrated Kerr micro-resonators can lead to highly nonlinear and complex optical spectra including frequency combs, instabilities, solitons and more. Due to the inherent nonlinearity, bistability and hysteresis in the system, the mapping of desired optical spectral properties to input parameters can be very difficult. Unlike previous approaches using traditional optimization or non-generative neural networks which struggle with multi-solution landscapes, we frame this inverse design problem as a generative distribution-learning task for the first time. Using conditional variational autoencoders and flow-matching models, we generate input parameters and their spectra for high-bandwidth steady-state solitons trained on Lugiato-Lefever equation simulations. Our approach can be applied to experimental data with little modification. Real experimental conditions for Kerr combs often deviate from theoretical models, making data-driven machine learning approaches particularly promising for applications in spectroscopy, optical communications, and nonlinear optics research.