Deep-Learning-Enabled State Classification in Thermally Stabilized Soliton Microcombs

Dissipative Kerr soliton (DKS) microcombs generated in microresonators feature octave-spanning bandwidths and gigahertz-to-terahertz repetition rates, underpinning applications from coherent communications to precision metrology. Reliable deployment, however, requires not only thermally robust soliton generation in chip-scale resonators but also real-time, noise-tolerant recognition of distinct soliton states beyond the limitations of heuristic methods and conventional Fourier analysis. Here, we present a hybrid deep-learning framework that integrates convolutional neural networks (CNNs) and Transformer-based multihead attention (TMHA), enabling precise and automated classification of measured soliton spectra. Stable DKS generation with 250 nm optical bandwidth is demonstrated in high-Q silicon nitride (Si3N4) microring resonators (MRRs), aided by phase-modulated sideband stabilization to suppress thermal instabilities. Single-soliton, multi-soliton, and perfect soliton crystal (PSC) states are consistently produced and quantitatively characterized via an improved fast Fourier transform–genetic algorithm (FFT-GA) routine. The hybrid CNN-TMHA model achieves a classification accuracy of 98.70%, with the shortest average inference time (6.51 ms) and highest throughput (153.58 samples/s) among all evaluated models. The combination of thermally stabilized soliton generation and high-performance state recognition enables field-ready, on-device recognition that supports closed-loop control, thereby moving microcombs from laboratory demonstrations to deployable microcomb-enabled photonic systems.