Machine learning-assisted dispersion engineering for predicting ultra-flat soliton microcombs

Flat soliton microcombs hold great potential for revolutionizing optical communication and computing. However, their practical realization is often hindered by the challenge in precisely controlling the microresonator's dispersion properties. In this work, we propose a simple inverse design approach that leverages a fully-connected neural network (FCNN) to effectively tailor microresonator dispersion for enabling ultra-flat soliton microcombs. We generated a dataset of microresonator geometries and their associated integrated dispersion (D_int) based on lithium niobate-on-insulator photonic platforms. By rigorously training the FCNN, we accurately predicted the optimal microresonator parameters necessary to achieve near-zero D_int profiles. To validate this approach, we conducted high-fidelity microcomb simulations with the open-source pyLLE solver. The results confirm the feasibility of generating ultra-flat soliton microcombs with exceptionally low power variations of 0.01 dB and 0.18 dB across the 180–205 THz range (corresponding to a wavelength span exceeding 200 nm).