Phase computation for the finite-genus solutions to the focusing nonlinear Schr & ouml;dinger equation using convolutional neural networks

We develop a method for retrieving a set of parameters of a quasi-periodic finite-genus (finite-gap) solution to the focusing nonlinear Schrodinger (NLS) equation, given the solution evaluated on a finite spatial interval for a fixed time. These parameters (named phases) enter the jump matrices in the Riemann-Hilbert (RH) problem representation of finite-genus solutions. First, we detail the existing theory for retrieving the phases for periodic finite-genus solutions. Then, we introduce our method applicable to the quasi-periodic solutions. The method is based on utilizing convolutional neural networks optimized by means of the Bayesian optimization technique to identify the best set of network hyperparameters. To train the neural network, we use the discrete datasets obtained in an inverse manner: for a fixed main spectrum (the endpoints of arcs constituting the contour for the associated RH problem) and a random set of modal phases, we generate the corresponding discretized profile in space via the solution of the RH problem, and these resulting pairs - the phase set and the corresponding discretized solution in a finite interval of space domain - are then employed in training. The method's functionality is then verified on an independent dataset, demonstrating our method's satisfactory performance and generalization ability.

Automated summary

We develop a method for retrieving the parameters of a quasi-periodic finite-genus solution to the nonlinear Schrodinger equation. The method utilizes convolutional neural networks optimized by Bayesian optimization to identify the best set of network hyperparameters.