Abstract:Physics--informed neural networks (PINN) have shown their potential in solving both direct and inverse problems of partial differential equations. In this paper, we introduce a PINN-based deep learning approach to reconstruct one-dimensional rough surfaces from field data illuminated by an electromagnetic incident wave. In the proposed algorithm, the rough surface is approximated by a neural network, with which the spatial derivatives of surface function can be obtained via automatic differentiation and then the scattered field can be calculated via the method of moments. The neural network is trained by minimizing the loss between the calculated and the observed field data. Furthermore, the proposed method is an unsupervised approach, independent of any surface data, rather only the field data is used. Both TE field (Dirichlet boundary condition) and TM field (Neumann boundary condition) are considered. Two types of field data are used here: full scattered field data and phaseless total field data. The performance of the method is verified by testing with Gaussian-correlated random rough surfaces. Numerical results demonstrate that the PINN-based method can recover rough surfaces with great accuracy and is robust with respect to a wide range of problem regimes.
What problem does this paper attempt to address?
This paper attempts to solve the problem of reconstructing one - dimensional rough surfaces from scattered electromagnetic field data. Specifically, the author introduced a deep - learning method based on Physics - Informed Neural Networks (PINN) for reconstructing the profiles of rough surfaces from scattered - field data under electromagnetic - wave irradiation.
### Problem Background
In the field of electromagnetic scattering, recovering the surface profile from the scattered electromagnetic field of a rough surface is a challenging problem, which is widely used in non - destructive testing, ground - penetrating radar, remote sensing, and biomedical imaging, etc. Traditional solutions include Rytov - type methods, Newton - type methods, sampling - reconstruction methods, and iterative methods, etc., but these methods usually rely on specific physical equations and have certain conditional limitations on the surface and incident wave (for example, the low - grazing - angle approximation).
### Solution Proposed in the Paper
To solve the above problems, the author proposed a new deep - learning method based on PINN. The main features of this method are as follows:
1. **Unsupervised Learning**: It does not require any prior surface data and is trained only with scattered - field data.
2. **Automatic Differentiation**: The surface function is approximated by a neural network, and the spatial derivative of the surface function is obtained by using automatic - differentiation techniques.
3. **Combination of Integral Equation and Method of Moments**: The scattered field is calculated using the Method of Moments (MOM), combined with the surface and its derivative predicted by the neural network.
4. **Loss - Function Design**: The neural network is trained by minimizing the mean - square error (MSE) between the predicted scattered field and the observed scattered field and adding boundary - condition constraints.
### Method Validation
The author verified the effectiveness and robustness of this method through numerical experiments. The experimental results show that the PINN - based method can recover rough surfaces with high precision and performs well under different noise levels, surface scales, surface heights, wave numbers, and incident angles.
### Formula Summary
- **Helmholtz Equation**:
\[
\nabla^2\psi(x,z)+k^2\psi(x,z) = 0
\]
where \(k=\frac{2\pi}{\lambda}\), and \(\lambda\) is the wavelength.
- **Kirchhoff - Helmholtz Integral Equation**:
\[
\psi(r)=\psi_i(r)+\int_S\mathbf{n}'\cdot\left[\psi(r)\nabla'G(r;r')-G(r;r')\nabla'\psi(r')\right]ds'
\]
where \(G(r;r')=\frac{i}{4}H_0^{(1)}(k|r - r'|)\), and \(H_0^{(1)}\) is the first - kind zero - order Hankel function.
- **Loss Function**:
\[
L(\theta,t):=L_{\text{eqn}}+L_{\text{bc}}
\]
where \(L_{\text{eqn}}\) is the mean - square error of the scattered field or the total field, and \(L_{\text{bc}}\) is the mean - square error of the boundary condition.
Through this method, the paper successfully solved the problem of reconstructing rough surfaces from scattered electromagnetic fields and demonstrated its potential in multiple application scenarios.