Modified failproof physics-informed neural network framework for fast and accurate optical fiber transmission link...

Area of Science:

• Scientific Machine Learning
• Optical Fiber Communications
• Nonlinear Optics

Background:

• Physics-informed neural networks (PINNs) are a powerful tool in scientific machine learning.
• Baseline PINNs face limitations in complex optical fiber communication modeling due to their loss function's non-convex landscape.
• These limitations are particularly evident in modeling soliton dynamics and pulse development in specialized fibers.

Purpose of the Study:

• To address the failure modes of baseline PINNs in complex optical fiber modeling.
• To enhance the accuracy and robustness of PINNs for simulating nonlinear phenomena like soliton propagation.
• To investigate the fundamental reasons behind PINN performance limitations in intricate scenarios.

Main Methods:

• Implementation of the scaffolding technique for PINN modeling.
• Application of the progressive block learning strategy for PINN modeling.
• Solving the nonlinear Schrödinger equation (NLSE) to model optical pulse propagation.

Main Results:

• The proposed methods significantly reduce errors in PINN-based optical fiber modeling.
• Accuracy increased by two to three orders of magnitude for complex modeling tasks.
• The study confirmed that performance issues stem from PINN design, not network architecture.

Conclusions:

• Scaffolding and progressive block learning effectively circumvent limitations of physics-based regularization in PINNs.
• These techniques enable more accurate modeling of optical pulse evolution dynamics.
• The findings offer a pathway to more reliable and accurate PINN applications in advanced optical systems.