Physics-based deep learning for modeling nonlinear pulse propagation in optical fibers.

Area of Science:

• Physics
• Optical Engineering
• Computational Science

Background:

• Nonlinear pulse propagation in optical fibers is crucial for modern communication systems.
• Traditional deep learning (DL) methods for modeling these phenomena often require extensive labeled data.
• Existing models struggle with accurately predicting complex nonlinear dynamics without ground truth.

Purpose of the Study:

• To introduce Phynet, a physics-based deep learning method for modeling nonlinear pulse propagation.
• To demonstrate Phynet's ability to learn and predict fiber dynamics independent of ground truth data.
• To overcome the data dependency limitations of conventional DL approaches.

Main Methods:

• Developed Phynet by integrating a handcrafted neural network with the nonlinear Schrödinger physics model.
• Optimized Phynet using a physics loss derived from the interaction between the neural network and the physical model.
• Utilized the inverse pulse propagation problem to evaluate Phynet's performance against a typical DL method.

Main Results:

• Phynet accurately restored initial pulse profiles with varying widths and powers.
• The method achieved prediction accuracy comparable to traditional DL techniques.
• Phynet demonstrated effective modeling of nonlinear fiber dynamics without relying on supervised loss or ground truth.

Conclusions:

• Phynet offers a viable alternative to data-hungry DL methods for modeling nonlinear fiber optics.
• The physics-informed approach reduces the need for extensive labeled datasets in training.
• This work provides new insights into modeling and predicting nonlinear dynamics in optical fibers.