Few-Shot Learning for Modulation Format Identification and Optical Performance Monitoring: A Domain-Adapted Approach

To address the limited generalization capability and heavy data dependency of conventional neural networks in dynamic elastic optical networks (EONs), a domain-adapted multi-task model-agnostic meta-learning (MT-MAML) framework is proposed for simultaneous modulation format identification (MFI) and optical signal-to-noise ratio (OSNR) estimation. By designating MFI as an auxiliary task to guide and constrain the OSNR regression task, MT-MAML employs a domain-agnostic initialization approach to optimize global meta-parameters, enabling rapid task adaptation under sparse-data conditions. Polar-coordinate constellation distributions were used as feature representations, which effectively capture the unique density and structural variations in both non-probabilistically shaped (non-PS) and probabilistically shaped (PS) QAM signals. A 32-Gbaud optical transmission system was established, and the simulation results were analyzed. The results demonstrate that when the sample length is set to 6000 under 5-shot setting, the proposed algorithm achieves 100% MFI accuracy together with an OSNR estimation MAE of about 0.2 dB. Furthermore, ablation experiments were constructed and the results confirm that MT-MAML significantly outperforms multi-task learning (MTL) and single-task model-agnostic meta-learning (ST-MAML) in convergence speed, estimation stability, and tolerance to transmission impairments. Specifically, MT-MAML maintains stable performance under residual chromatic dispersion (RCD) ranging from −600 ps/nm to +600 ps/nm, and its OSNR MAE increases only marginally (from about 0.21 dB to 0.25 dB) across extended transmission distances from 200 km to 1000 km and launch powers from 0 dBm to 6 dBm. These results highlight the strong environmental robustness and deployment potential of the proposed scheme for dynamic elastic optical networks.