Optimization of physics-informed neural networks for nonlinear Schrödinger equation

In this paper, PINN is applied to the NLSE equation in order to determine the performance range and limiting factors. Some tools, such as manual weights of the loss function components, and selective application of the sinusoidal activation function, are applied to improve the results. Accepting the fact that PINN loses to SSFM in terms of performance, the application of Meta-PINN to NLSE is investigated to cover the range of parameters, demonstrating the successful generalisation ability of Meta- PINN. The paper concludes with a recommendation on how to tune PINN to successfully solve NLSE.