Data-driven vector soliton solutions of coupled nonlinear Schrödinger equation using a deep learning algorithm (Q2246919)

The aim of the paper is to investigate data-driven vector solitons for the coupled nonlinear Schrödinger equation (CNLSE) by means of an improved physics-informed neural networks (PINN) method. One considers an initial and boundary Dirichlet problem for the CNLSE: \[ iq_t + \frac{1}{2}q_{xx}+{\|q\|} _2^2q =0,\text{ for }x\in [- x_L,x_L], t\in[t_0,t_T], \] \[ q(x,t_0)=q^{t_0}(x),\text{ for }x\in [- x_L,x_L], \] \[ q(x_L,t)=q(-x_L,t)=q^b(t),\text{ for }t\in[t_0,t_T] \] Here, $q(x,t)=(q_1(x,t),q_2(x,t))$ is a complex vector field. The original PINN method for solving CNLSE is briefly reviewed and in order to improve the convergence rate and the approximation ability of the original PINN method a new pre-fixed multistage training algorithm is proposed. In the second section of the article the construction of an improved PINN by pre-fixed multi-stage training is largely described, the flow diagram together with the pre-fixed multi-stage training algorithm are shown. Few words on the efficiency and accuracy of the method end the section. In the third section of the article the performance of the improved method is investigated. One considers the case of a nondegenerate vector solution soliton and the collision of vector soliton solution. References contain 65 titles.