Abstract

As global data traffic continues to surge, more spectrum resources are desired for optical networks to avoid capacity crunch. Among multiple solutions, multi-band transmission is appealing for practical deployment by utilizing the existing fiber infrastructures. In a multi-band system, multi-pump Raman amplifiers (RAs) are an attractive option for wide-band amplification with low noise figures. The physical characteristics of an RA can be described as a set of nonlinear ordinary differential equations, of which there is no closed-form solution and the calculation is cumbersome. To efficiently shape the gain profile of an RA, many previous works utilize neural networks (NNs) to build an offline-trained model as the digital twin (DT) for the RA. Then, optimization algorithms can be applied to the NN-based DT. However, this type of method requires a large training dataset containing various pump configurations and corresponding gain profiles, resulting in high complexity during the offline training stage. In this paper, we propose a novel scheme called SMOF, which conducts RA modeling and gain profile optimization simultaneously. By iteratively freezing and unfreezing the inner parameters of the DT, the pump configuration is adjusted to achieve a desirable gain profile. A real-time experiment with a 4-pump RA for C+L-band is conducted for performance evaluation. SMOF can optimize one gain profile within ∼10 iterations of pump reconfigurations and optimize multiple gain profiles within ∼30 iterations of pump reconfigurations, thus achieving data-efficient optimization. Additionally, the DT obtained during optimization shows an optimal accuracy around the optimization target, which can be utilized to assist network control and optimization.