Distributed nonlinear power dynamics prediction across 40 channels in a wideband discrete Raman amplifier using a mixture-of-experts framework

In this work, we demonstrate the distributed optical power evolution of a wideband discrete Raman amplifier in a single-mode fiber across 40 channels. Our approach predicts how the optical power distributes along the entire optical fiber length with spatial resolution measured in kilometers for all 40 channels simultaneously. To achieve this, we employ a deep learning network based on a mixture-of-experts (MoE) framework. This marks a clear advance over other neural network-based approaches found in the scientific literature, which usually provide only pointwise or end-of-fiber estimates or rely on computationally demanding numerical solvers of the governing differential equations. In our case, the proposed MoE-based framework predicts the power propagation along the fiber at a 1 km spatial resolution across multiple channels. All aspects of the discrete Raman amplifier modeling, including signal and pump power evolution, were entirely developed and implemented within the MATLAB environment. Furthermore, we combined the MoE model with the Optuna multi-objective optimizer as a proof of concept to demonstrate its applicability in Raman amplifier optimization tasks. To our knowledge, this is the first integration between a MoE framework and Optuna in Raman amplifiers, highlighting a novel methodological contribution. Specifically, this approach achieves consistent performance, delivering a root mean squared error (RMSE) value below 0.02 for all 40 channels of the Raman amplifier while remaining computationally efficient.