Abstract:We present a novel deep learning method for computing eigenvalues of the fractional Schrödinger operator. Our approach combines a newly developed loss function with an innovative neural network architecture that incorporates prior knowledge of the problem. These improvements enable our method to handle both high-dimensional problems and problems posed on irregular bounded domains. We successfully compute up to the first 30 eigenvalues for various fractional Schrödinger operators. As an application, we share a conjecture to the fractional order isospectral problem that has not yet been studied.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is the eigenvalue problem of calculating the fractional - order Schrödinger operator. Specifically, the author proposes a new deep - learning method to calculate the eigenvalues of the fractional - order Schrödinger operator in a bounded region. The following are the main problems and background of the paper:
### Research Background
1. **Applications of Fractional - order Partial Differential Equations**:
- Fractional - order partial differential equations have shown effectiveness in modeling abnormal diffusion phenomena, turbulence, porous media flow, etc.
- In quantum mechanics, Laskin introduced the fractional - order Schrödinger equation by using Levy flights instead of classical Brownian motion, revealing some new phenomena that cannot be explained by the standard Schrödinger equation.
2. **Challenges of Numerical Methods**:
- Due to the non - locality and singularity of fractional - order partial differential equations, most numerical methods mainly focus on one - dimensional problems.
- For two - dimensional problems, the commonly used method is the adaptive finite element method, but the application of these methods in irregular domains faces significant challenges.
### Paper Objectives
1. **Definition of the Eigenvalue Problem**:
- Study the eigenvalue problem of the fractional - order Schrödinger operator in a bounded region \(\Omega\subset\mathbb{R}^d\):
\[
\begin{cases}
(-\Delta)^s u(x)+V(x)u(x)=\lambda u(x), & x\in\Omega, \\
u(x) = 0, & x\in\Omega^c\equiv\mathbb{R}^d\setminus\Omega.
\end{cases}
\]
- Here, \(V(x)\) is a real - valued potential function, and \((-\Delta)^s\) is the fractional - order Laplacian operator, defined as:
\[
(-\Delta)^s u(x)=C_{d,s}\int_{\mathbb{R}^d}\frac{u(x)-u(y)}{\|y - x\|^{d + 2s}}\,dy,
\]
where \(C_{d,s}=\frac{2^{2s}s\Gamma(s + d/2)}{\pi^{d/2}\Gamma(1 - s)}\).
2. **Limitations of Existing Methods**:
- At present, the exact eigenvalues are unknown, and there are no analytical formulas to calculate them.
- Although numerical methods such as spectral methods, finite difference methods, and finite element methods can approximately calculate eigenvalues, they perform poorly in high - dimensional and irregular domains.
### Proposed Method
1. **Deep - learning Method**:
- Transform the original eigenvalue problem into a series of minimization problems with constraints.
- Introduce a new loss function, including a penalty term to efficiently handle orthogonal constraints.
- Design a new neural network architecture, combining the eigenfunctions in prior knowledge, especially singularity and boundary conditions.
2. **Numerical Experiments**:
- Evaluate the accuracy of the method by solving a series of problems that can be processed by spectral methods.
- Calculate problems from 1 to 9 dimensions with a relative error of less than 0.1%.
- Further calculate up to 30 eigenvalues, with the error remaining below 1%.
- Test the method in general domains and compare the results with those of the finite element method, showing higher accuracy.
### Applications
1. **Fractional - order Isospectral Problems**:
- Apply the method to fractional - order isospectral problems and propose a conjecture based on numerical results: even if the spectra of two domains are the same under the Laplacian operator, their spectra will not be the same under the fractional - order Laplacian operator.
### Conclusion
The paper proposes a new deep - learning method that can calculate the eigenvalues of the fractional - order Schrödinger operator efficiently and accurately, especially performing well in high - dimensional and irregular domains. This method is not only innovative in theory but also has high practical value in practical applications.