Image classification using spectral features from a Kerr comb photonic reservoir computer

In this work, we present a photonic classification architecture based on a Kerr microresonator operating as a nonlinear reservoir. The system exploits chaotic optical frequency comb formation in a high-Q microresonator to transform time-encoded input signals into high-dimensional spectral representations. Using the QR-compressed MNIST handwritten digit dataset, we numerically demonstrate that the proposed architecture improves classification accuracy from 71.67% to 83.88% compared with a purely linear baseline, even when only a small subset of input features is used. We further analyze the influence of key resonator parameters on classification performance. These results highlight the potential of nonlinear microresonator-based frequency comb reservoirs as compact photonic processors for machine-learning tasks with low footprint and simple readout.