Data-driven nondegenerate bound-state solitons of multicomponent Bose–Einstein condensates via mix-training PINN

Abstract

In this paper, by modifying loss function MSE and training area of the physics-informed neural network (PINN), we proposed two neural network models: mix-training PINN and prior information mix-training PINN. We demonstrated the advantages of these models by simulating nondegenerate bound-state solitons (NDBSSs) of multicomponent Bose–Einstein condensates (BECs). Numerical experiments showed that our proposed models are not only simulate the NDBSSs of multicomponent BECs, but also significantly improve the simulation capability. Compared with original PINN, the prediction accuracy of our proposed models are improved by one to three orders of magnitude. By testing the inverse problem of multicomponent BECs, it is also proved that these models have good performance.

Keywords

• Physics-informed neural network
• Mix-training
• Prior information
• Nondegenerate bound-state solitons
• Multicomponent Bose–Einstein condensates