🔬 optics

Physics-Informed Neural Networks for 2D Plane Wave Scattering in Arbitrary Dielectric Structures

This paper introduces a meshless physics-informed neural network framework that solves 2D electromagnetic wave scattering in arbitrary dielectric structures by embedding Maxwell's equations into the loss function, achieving high accuracy for TM polarization and overcoming TE polarization singularities through hyperbolic-tangent smoothing at dielectric boundaries.

Original authors: Zheng-Yu Huang, Yu Tian, Jing-Wen Zhang, Nicolae C. Panoiu

Published 2026-07-31

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Zheng-Yu Huang, Yu Tian, Jing-Wen Zhang, Nicolae C. Panoiu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict how a ripple moves when it hits a rock in a pond. In the world of physics, light behaves a bit like that ripple, but instead of water, it's an invisible dance of electric and magnetic fields. When this light hits an object—like a glass bead or a metal chip—it scatters, bouncing off in all sorts of directions. Scientists have spent decades trying to calculate exactly how this scattering happens using giant, complex grids of numbers, kind of like drawing a massive checkerboard over the pond to track every single drop of water. But these grids are heavy, slow, and sometimes break when the shapes get too weird or the materials too tricky.

Enter a new kind of detective: the Physics-Informed Neural Network (PINN). Think of a standard computer program as a student who memorizes a textbook but doesn't understand the rules of the game. A PINN, however, is like a student who is given the rulebook (the laws of physics) and told to solve the puzzle by figuring out the answer that fits those rules perfectly, without needing to draw a checkerboard. Instead of a grid, it uses a flexible, mesh-free brain that learns the shape of the light waves directly. This paper dives into how well this "rule-following brain" can handle the messy, jagged world of light hitting strange, dielectric objects (materials that let light pass through but slow it down, like glass or plastic), specifically testing if it can handle the two main ways light can wiggle: side-to-side (Transverse Magnetic) and up-and-down (Transverse Electric).

The Paper's Story: Teaching a Brain to See Light

The researchers in this paper set out to build a super-smart, mesh-free framework to solve a classic problem: how does light scatter when it hits 2D objects made of different materials? They didn't just guess; they built a neural network—a type of artificial intelligence—and forced it to obey the strict laws of electromagnetism. They embedded the famous Maxwell equations (the rulebook for how electricity and magnetism behave) directly into the network's "loss function." In plain English, this means the AI gets a "punishment score" every time it predicts a light wave that breaks the laws of physics. The goal is to minimize this score until the AI's prediction is perfectly legal.

They tested this AI on four different scenarios, ranging from a single glass cylinder to a chaotic cluster of three cylinders and even a weird mix of a square, a triangle, and a circle. They wanted to see if the AI could predict the scattered light fields for two different "modes" of light: Transverse Magnetic (TM) and Transverse Electric (TE).

The Good News: The TM Mode
For the Transverse Magnetic (TM) polarization, the AI was a star. It predicted the scattered light fields with incredible accuracy. When they compared the AI's results to the known mathematical answers for single and double cylinders, the errors were tiny—mostly staying below 0.1 (or 10%). Even when they threw three cylinders at it, the AI handled the complex interactions smoothly. The paper suggests that for these shapes, the AI learned the physics so well that it could predict the light's behavior almost perfectly, matching the results of traditional, heavy-duty simulation methods.

The Bad News: The TE Mode and the "Sharp Edge" Problem
However, things got tricky with the Transverse Electric (TE) polarization. Here, the AI stumbled. The paper explains that when light hits a sharp boundary between two materials (like the edge of a glass cylinder), the math for the TE mode gets "singular." Imagine trying to draw a smooth curve that suddenly has to jump up and down infinitely fast at a specific point; a neural network, which prefers smooth, gentle curves, gets confused and starts making mistakes. The errors for the TE mode were huge—sometimes nearly 100% off—because the AI couldn't handle the sudden, jagged changes in the material properties.

The Fix: The "Softening" Trick
The researchers didn't give up; they found a clever workaround. They realized the problem was the "sharpness" of the boundary in the math. So, they introduced a "smoothing function" using a hyperbolic tangent curve. Think of this as taking a razor-sharp knife edge and sanding it down just a tiny bit to make it a gentle slope. By telling the AI that the material properties change gradually over a tiny distance (about $10^{-3}$ micrometers, which is in the nanometer range) instead of jumping instantly, the math stopped having those impossible singularities.

When they applied this "sanding" trick, the AI's performance for the TE mode skyrocketed. The errors dropped significantly, and the predictions began to match the benchmark solutions (calculated using the traditional Finite-Difference Time-Domain method) very closely.

What They Found and What They Didn't
The paper concludes that this mesh-free, physics-informed approach is robust and scalable. It works beautifully for TM polarization across all tested shapes. For TE polarization, it works just as well if you use the smoothing trick to handle the sharp edges. The authors explicitly note that without this smoothing, the method fails for TE modes in complex structures. They also confirm that the method converges stably, meaning the AI doesn't get stuck or go crazy during training; it steadily learns the solution.

The study was conducted through numerical simulations, not physical experiments with lasers and glass. They tested single cylinders, concentric shells, three-cylinder clusters, and irregular shapes. The results show that while the AI is a powerful tool for predicting light scattering, it needs a little help (the smoothing function) to handle the most difficult, jagged mathematical corners of the TE mode. The paper suggests this framework is a promising step toward solving complex electromagnetic problems without the need for massive, rigid grids, but it leaves the door open for future work on 3D shapes and even more complex materials.

Technical Summary: Physics-Informed Neural Networks for 2D Plane Wave Scattering in Arbitrary Dielectric Structures

Problem Statement
Numerical investigation of electromagnetic wave scattering is critical for applications in optics, radar, and materials science. Traditional methods such as the Finite-Difference Time-Domain (FDTD), Finite Element Method (FEM), and Boundary Element Method (BEM) rely on mesh discretization, which can limit scalability and geometric flexibility, particularly for multiscale problems and structures with highly irregular boundaries. While Physics-Informed Neural Networks (PINNs) offer a meshless alternative by embedding physical laws into the loss function, their application to high-frequency electromagnetic scattering faces specific challenges. These include accuracy loss due to automatic differentiation at strongly discontinuous dielectric interfaces, high-frequency oscillations, and convergence difficulties. Furthermore, existing approaches often struggle to handle the distinct mathematical singularities arising in Transverse-Electric (TE) polarization compared to Transverse-Magnetic (TM) polarization, and a unified framework for stable, reliable results across complex configurations remains an open challenge.

Methodology
The authors propose a meshless computational framework based on PINNs to solve 2D electromagnetic scattering in inhomogeneous media. The core of the method involves using a feedforward neural network (FNN) with hyperbolic tangent activation functions to approximate the continuous physical fields (specifically the scattered field components $E_z$ for TM and $H_z$ for TE).

The framework embeds the frequency-domain Maxwell equations and radiation boundary conditions directly into the network's loss function. The total loss function, $L(\theta)$ , is a weighted sum of:

Physics Residual ( $L_{PDE}$ ): The mean squared error of the residuals of the governing partial differential equations (PDEs) computed via automatic differentiation at internal collocation points.
Boundary Residual ( $L_{BC}$ ): The error enforcing radiation boundary conditions (approximated as first-order absorbing conditions) at the domain boundaries.

The governing equations are derived by separating the total field into incident and scattered components. For TM polarization, the problem reduces to a 2D Helmholtz equation. For TE polarization, the governing equation involves a divergence operator acting on the inverse permittivity ( $1/\varepsilon_r$ ), which introduces mathematical singularities at dielectric interfaces where $\varepsilon_r$ is discontinuous.

To address the convergence issues in TE polarization caused by these singularities, the authors introduce a hyperbolic-tangent smoothing function at dielectric boundaries. This function creates a transition region of width $\delta$ , smoothing the discontinuous permittivity profile to remove the singularities associated with the derivatives of $1/\varepsilon_r$ .

Key Contributions

Unified Framework: The paper establishes a single PINN-based framework capable of handling both TM and TE polarizations for arbitrary 2D dielectric structures without explicit mesh generation.
Handling TE Singularities: The authors identify that the standard PINN approach fails to achieve high accuracy for TE polarization due to singularities in the governing equations at material interfaces. They propose and validate a smoothing technique using a hyperbolic tangent function to regularize the permittivity distribution, significantly improving accuracy.
Comprehensive Validation: The method is tested across four distinct configurations: a single cylinder, concentric multilayer cylindrical shells, three arbitrarily arranged cylinders, and composite irregular structures (square, cylinder, triangle).
Performance Analysis: The study provides a detailed analysis of convergence behavior, loss function evolution, and the impact of transition region width and sampling density on the relative $L_2$ error.

Results

TM Polarization: The method demonstrates high accuracy for TM polarization across all tested configurations. The relative $L_2$ errors remain consistently low (mostly $\le 0.1$ ), even for complex multi-cylinder and irregular structures. The results closely match analytical solutions and FDTD benchmarks.
TE Polarization (Unsmoothed): Without modification, the method exhibits significantly higher errors for TE polarization (e.g., relative $L_2$ error of 0.9184 for a single cylinder), attributed to the singularities in the governing equations at sharp dielectric boundaries.
TE Polarization (Smoothed): Introducing the hyperbolic-tangent smoothing function drastically improves TE results. The errors decrease substantially, and the predictions align closely with FDTD benchmarks. The study finds that a transition width of $\delta \approx 5 \times 10^{-3}$ (in micrometer units) offers an optimal balance between stability and accuracy.
Convergence: The training process exhibits stable convergence across all configurations. The loss function, gradient norms, and step-size norms decline steadily without oscillation or divergence, confirming the robustness of the optimization strategy (using Adam and L-BFGS optimizers with dynamic weighting).

Significance and Claims
The paper claims that the proposed framework offers a robust and scalable solution for complex electromagnetic scattering problems. It demonstrates that PINNs can effectively predict scattered fields for both polarizations in arbitrary dielectric structures, provided that the specific mathematical challenges of TE polarization (interface singularities) are addressed via smoothing. The authors conclude that their meshless approach successfully overcomes the limitations of traditional discretization methods regarding geometric flexibility and computational cost, while maintaining physical consistency through the direct embedding of Maxwell's equations. The work validates the potential of PINNs as a versatile tool for modeling wave interactions in micro/nanostructured materials, though it acknowledges that future work is needed to extend these capabilities to three-dimensional geometries and advanced material models.

More like this