PINN deep learning method for the Chen–Lee–Liu equation: Rogue wave on the periodic background

Wei‐Qi Peng, Juncai Pu, Yong ChenCommunications 2021 · DOI · arXiv

Introduction

The Chen-Lee-Liu (CLL) equation, a derivative nonlinear Schrödinger equation, serves as a cornerstone in modeling nonlinear optics and plasma physics. It specifically describes the propagation of self-steepening optical pulses in fibers and plasma waves in astrophysical contexts. While analytical solutions for simple solitons exist, "rogue waves"—massive, spontaneous spikes—are notoriously difficult to model, particularly when they manifest against a non-flat, periodic background. Traditionally, these solutions require complex mathematical frameworks involving Jacobian elliptic functions.

Recent research has pivoted toward Physics-Informed Neural Networks (PINNs) to solve these complex dynamics numerically. By embedding the underlying physical laws directly into the neural network's architecture, researchers can bypass traditional mathematical "heavy lifting" and discover solutions with minimal data.

Methodology: Embedding Physics into Deep Learning

The core of this approach lies in the PINN framework. Unlike standard deep learning models that rely solely on data, PINNs use the governing partial differential equation (PDE) as a regularizer. For the complex-valued CLL equation, the authors decomposed the system into real ($u$) and imaginary ($v$) components, resulting in two coupled real-valued equations.

The network architecture typically consists of a multi-layer perceptron with 9 hidden layers and 40–60 neurons per layer, utilizing the hyperbolic tangent (tanh) activation function. Optimization is performed using the L-BFGS algorithm, a second-order method that excels in navigating the smooth loss landscapes characteristic of physical systems.

Training data is sparse, often consisting of only 100 to 400 points sampled via Latin Hypercube Sampling (LHS) from initial and boundary conditions. However, the model is constrained by 10,000 to 20,000 "collocation points" where the network must satisfy the CLL equation, ensuring the output remains physically consistent.

Capturing Complex Wave Dynamics

The PINN successfully modeled several wave types, starting with standard periodic waves. These repeating patterns serve as the baseline for more complex interactions.

Numerical simulation of a periodic wave solution for the CLL equation using PINNs.

The most significant achievement, however, was the accurate modeling of rogue periodic waves. These are extreme events where a high-amplitude spike emerges spontaneously from a repeating background pattern.

The standout result: A rogue wave emerging from a periodic background, captured via deep learning.

The PINN achieved an $L_2$ relative error of $6.10 \times 10^{-3}$ for these rogue periodic waves. Interestingly, the training process for these waves exhibited a "staircase" loss curve, suggesting that the network first masters the periodic background before identifying the localized, high-energy rogue spike.

Advantages and Future Directions

PINNs offer a "mesh-free" alternative to traditional Finite Difference Methods (FDM). They can predict the state of a wave at any coordinate $(x, t)$ without being tethered to a predefined grid, providing greater flexibility in studying extreme events.

Despite these successes, the model faces challenges with temporal scaling. Performance tends to degrade over long simulation windows as the network struggles to maintain physical consistency. Future research may look toward reservoir computing or using the Lax equation as a more stable physical constraint to extend the temporal reach of these deep learning models.

Summary updated

BibTeX
@inproceedings{peng2021pinn,
  author = {Wei‐Qi Peng and Juncai Pu and Yong Chen},
  title = {{PINN deep learning method for the Chen–Lee–Liu equation: Rogue wave on the periodic background}},
  booktitle = {Communications},
  year = {2021},
  doi = {10.1016/j.cnsns.2021.106067},
}
Figure