Abstract:We utilize the Feature Decoupling Distributed (FDD) method to enhance the capability of deep learning to fit the Nonlinear Schrodinger Equation (NLSE), significantly reducing the NLSE loss compared to non decoupling model.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is how to improve the accuracy of deep - learning models when fitting the Nonlinear Schrödinger Equation (NLSE). Specifically, the authors use the Feature Decoupling Distributed (FDD) method to enhance the fitting ability of deep - learning models to NLSE, thereby significantly reducing NLSE loss and improving the generalization ability of the model under different transmission distances and boundary conditions.
### Problem Background
1. **Importance of NLSE**:
- NLSE describes the propagation of optical pulses in fiber channels and is the basis for studying nonlinear optics and fiber - optic communications.
- Traditional methods for solving NLSE are mainly based on the Split - Step Fourier Method (SSFM), but this method has high computational complexity and is difficult to meet the needs of practical engineering applications.
2. **Application of Deep Learning**:
- Deep Learning (DL), due to its powerful nonlinear fitting ability and parallel computing efficiency, has become an effective method for solving NLSE and modeling fiber channels in recent years.
- However, most existing DL schemes are either purely data - driven or introduce physical prior knowledge in the loss function, causing neural networks to need to learn the linear and nonlinear coupling characteristics of the entire NLSE system, which will lead to a decrease in accuracy in more complex scenarios.
### Solution
To overcome the above problems, the authors propose the FDD method:
1. **Feature Decoupling**:
- The FDD method decouples linear features from nonlinear features by combining prior physical models.
- This method enables neural networks to focus on optimizing the nonlinear part, thereby reducing the fitting difficulty.
2. **Linear System Modeling**:
- A linear system modeled by physical formulas is cascaded at the back of the neural network to compensate for dispersion and attenuation effects.
- This can ensure that the overall system can better fit the complete NLSE.
3. **Improved Training Strategy**:
- By introducing the encoded input method, the transmission distance parameter \( z \) is encoded into the input end, enhancing the model's generalization ability for different distances.
- The derivatives of the output signal with respect to the distance \( z \) and time \( t \) are calculated to evaluate NLSE loss.
### Experimental Results
1. **Waveform Prediction Accuracy**:
- In training and non - training scenarios, the Normalized Mean Square Error (NMSE) of the FDD model is reduced by more than three times and ten times respectively.
- For transmission distances of 10 - 100 kilometers, the NMSE accuracy of the FDD model remains at a low level, and the NLSE loss is reduced by two orders of magnitude.
2. **Generalization Ability**:
- The FDD model shows higher prediction accuracy at different transmission distances, indicating that it has stronger distance generalization ability and the ability to adapt to a wider range of boundary conditions.
### Conclusion
By introducing the FDD method, the authors have successfully improved the accuracy and generalization ability of deep - learning models in fitting NLSE. This method is not only applicable to NLSE, but can also be extended to the modeling of other physical processes controlled by Partial Differential Equations (PDEs), enabling neural networks to better fit these equations.
### Formula Summary
1. **General form of NLSE**:
\[
\frac{\partial s}{\partial z}+\alpha s + i\beta_2\frac{\partial^2 s}{\partial t^2}-\frac{8}{9}i\gamma|s|^2 s = 0
\]
where \( s = s(z, t) \) represents the optical field envelope, \( z \) is the propagation distance, and \( t \) is the time, and \( \beta_2 \) is the dispersion coefficient, and \( \alpha \)