Task-aware meta equalizer for multi-scenario generalization in coherent DWDM systems

In dense wavelength division multiplexing (DWDM) systems, neural networks have been widely employed to compensate for both linear and nonlinear channel distortions. However, current normal neural network equalizers (NNE) exhibit limited generalization when encountering variations in modulation formats, transmission rates, and optical signal-to-noise ratio (OSNR) conditions. Retraining from scratch in each new scenario is resource-intensive, rendering these models impractical for the highly dynamic configurations of elastic optical networks. To address this challenge, we propose a task-aware neural network equalizer based on Meta-SGD, which models different channel configurations as independent tasks within a multi-task learning framework. During the test phase, we adopted a two-stage training strategy, combining meta-learning for rapid adaptation with transfer learning for fine-tuning under limited data conditions. Experiments were conducted on a large-scale DWDM dataset spanning the S, C, and L bands, covering 263 channels, a 200 km transmission distance, and diverse combinations of modulation formats and transmission parameters. The results demonstrate that the proposed approach significantly outperforms NNE methods on both seen and unseen tasks. With only 20% of a set of data, the Meta-SGD model achieves a higher Q factor after just 10 steps of fast adaptation and converges within 10 epochs on seen tasks and 20 epochs on unseen tasks. In contrast, NNE models require up to 70% of a set of data and over 50 training epochs to reach comparable performance. Moreover, when trained with only 20% of the data, the NNE models fail to reach optimal performance even after 100 epochs, highlighting the superior data efficiency and generalization capability of the proposed Meta-SGD-based approach.