Dual-End Physics-Based Neural Augmentation for Robust Fiber Impairment Compensation: Mitigating Pre- and Post-Model Errors

The performance of fiber-optic systems is limited by impairments such as dispersion and nonlinearities, which are mitigated by physics-based models (e.g., digital backpropagation) or data-driven methods. A hybrid cascaded learning (CL) approach improves performance by using neural networks (NNs) to correct post-model residual errors, but it is not robust to parameter mismatches that distort the model's input and induce pre-model errors. We propose a neural augmentation (NA) scheme where an NN preprocesses the model input to correct these pre-model errors, enhancing robustness. By combining NA and CL, we introduce a dual-end augmented network (DEAN) that concurrently addresses both error types. Implemented with reservoir computing, the DEAN architecture generates dual corrections with minimal computational overhead. Numerical analysis of a 32-Gbaud 16-QAM system demonstrates that DEAN reduces the bit error rate by an order of magnitude compared to CL or NA alone, while maintaining a similar level of computational complexity. This approach bridges the gap between model accuracy and adaptability, enabling reliable fiber communication under real-world parameter variations.