Surface profile recovery from electromagnetic field with physics-informed neural networks

Yuxuan Chen, Ce Wang, Yuan Hui, Mark SpivackarXiv 2024 · OpenReview

TL;DR

This paper introduces a novel approach using Physics-Informed Neural Networks (PINNs) to reconstruct the profile of rough surfaces from electromagnetic field data. By integrating physical laws (governing equations of wave scattering) directly into the neural network's training process, the method achieves accurate surface recovery without requiring prior knowledge of the surface itself, making it a powerful tool for real-world applications.

What Problem Does This Paper Address?

Recovering the shape of rough surfaces from scattered electromagnetic waves is a crucial problem in various fields, including non-destructive testing, geophysics, and remote sensing. Traditional methods for this "inverse scattering problem" often rely on iterative numerical techniques that can be sensitive to initial guesses and specific conditions, or require extensive datasets of known surface profiles and their corresponding scattered fields. The challenge lies in developing a robust and efficient method that can accurately reconstruct complex surface geometries from limited or noisy field measurements, ideally without needing ground truth surface data.

Methods

The core of this work is the application of Physics-Informed Neural Networks (PINNs) to the inverse scattering problem of rough surface reconstruction. Here's a breakdown of their approach:

  • Neural Network as Surface Representation: A deep neural network is used to approximate the unknown rough surface profile, $z = h(x)$. The network takes a spatial coordinate $x$ as input and outputs the corresponding height $h(x)$.
  • Automatic Differentiation for Derivatives: Crucially, the spatial derivatives of the surface height ($\partial h/\partial x$ and $\partial^2 h/\partial x^2$) are computed directly from the neural network's output using automatic differentiation. This is a key advantage of PINNs, as it avoids explicit parameterization of derivatives.
  • Physics-Informed Loss Function: The neural network is trained by minimizing a loss function that comprises two main components:
  • Scattered Field Loss ($L_{eqn}$): This term enforces the physics of wave scattering. The predicted surface and its derivatives are fed into a Method of Moments (MOM) solver, which calculates the scattered electromagnetic field. The loss is the mean squared error (MSE) between this calculated scattered field and the measured scattered field data.
  • Boundary Condition Loss ($L_{bc}$): This term incorporates known constraints on the surface height at specific boundary points, ensuring physically plausible solutions.
  • Unsupervised Learning: The method is unsupervised in the sense that it does not require a dataset of known surface profiles. It learns solely from the observed field data and the underlying physics equations.
  • Two Field Types: The approach is demonstrated for both Transverse Electric (TE) and Transverse Magnetic (TM) incident fields, corresponding to Dirichlet and Neumann boundary conditions, respectively.
  • Two Data Types: The method is tested with two types of field data: full scattered field data (including amplitude and phase) and phaseless total field data (amplitude only).

Datasets

The paper does not use pre-existing datasets for training. Instead, it generates synthetic data for testing and validation. The "datasets" are essentially simulated electromagnetic field measurements (both full scattered field and phaseless total field) obtained from synthetically generated Gaussian-correlated random rough surfaces. These surfaces are characterized by parameters like surface scale (autocorrelation length), peak-to-trough height, wavenumber, and angle of incidence. The method is then evaluated on its ability to reconstruct these known synthetic surfaces from the simulated field data.

Key Insights and Ideas

  • Physics as a Regularizer: By embedding the governing PDEs (via MOM and integral equations) into the loss function, PINNs act as a powerful regularizer. This allows the network to learn physically consistent surface profiles even with limited or noisy data, a significant advantage over purely data-driven methods.
  • Unsupervised Reconstruction is Feasible: The success of this PINN-based approach demonstrates that accurate surface reconstruction is possible without needing ground-truth surface data, which is often difficult and expensive to obtain in practice.
  • Automatic Differentiation is Key: The ability to obtain surface derivatives directly from the neural network via automatic differentiation is a critical enabler for integrating the physics of scattering into the learning process.
  • Robustness to Noise: The method shows remarkable robustness to noise in the measured field data, maintaining reasonable accuracy even with up to 15% noise in the full scattered field case.
  • Flexibility in Data Types: The PINN framework can effectively handle both full scattered field data and the more challenging phaseless total field data.

Key Results

  • High Accuracy Reconstruction: The PINN-based method successfully reconstructs one-dimensional Gaussian-correlated rough surfaces with high accuracy, closely matching the actual surface profiles in both TE and TM field cases.
  • Robustness to Noise: The method can tolerate noise levels up to approximately 15% for full scattered field data and 5% for phaseless total field data while still producing recognizable surface shapes.
  • Effective with Varying Parameters: The approach demonstrates robustness across a range of problem regimes, including variations in surface scale, peak-to-trough height, and incident wave characteristics.
  • Performance Degradation with Extreme Heights: While generally robust, reconstruction performance degrades for surfaces with very large peak-to-trough heights (greater than $\lambda$).
  • Error Increases with Wavenumber: Higher wavenumbers (frequencies) lead to increased error, suggesting that deeper neural networks might be needed for such scenarios.

Limitations and Future Work

  • Computational Cost of MOM: The paper notes that extending this method to 3D problems would significantly increase the computational complexity due to the inversion of larger matrices in the Method of Moments.
  • Error with High Wavenumbers: The error increases with higher wavenumbers, suggesting that deeper neural networks or other architectural improvements might be necessary for such cases.
  • Future Work: The authors suggest that the method can be extended to 3D problems and to inverse waveguide problems, which involve reconstructing pairs of rough surfaces. They also hint at the potential for treating real-world applications like non-destructive testing and microwave remote sensing.

Summary updated

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.
BibTeX
@misc{chen2024surface,
  author = {Yuxuan Chen and Ce Wang and Yuan Hui and Mark Spivack},
  title = {{Surface profile recovery from electromagnetic field with physics-informed neural networks}},
  year = {2024},
  eprint = {2404.14984},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2404.14984},
}
Figure