PINNs for Electromagnetic Wave Propagation

This study presents a hybrid Physics-Informed Neural Network (PINN) framework that overcomes traditional accuracy and energy conservation limitations in electromagnetic wave propagation by integrating time-marching, causality-aware weighting, interface continuity losses, and a Poynting-based regularizer, thereby achieving FDTD-comparable performance without labeled training data.

The Gist

Imagine you are trying to teach a computer to predict how ripples move across a pond. In the world of physics, these ripples are electromagnetic waves (like light or radio signals), and the "pond" is a box with metal walls (a cavity).

For decades, scientists have used a very strict, rule-following method called FDTD (Finite-Difference Time-Domain) to simulate this. Think of FDTD as a grid of tiny buckets. You pour water into one bucket, and the rules tell you exactly how much flows into the neighbors. It's accurate, but it's rigid and requires a lot of computing power to build the grid.

Then came PINNs (Physics-Informed Neural Networks). Think of a PINN not as a grid of buckets, but as a super-smart, shape-shifting artist. Instead of following a grid, the artist tries to "guess" the shape of the wave everywhere at once by learning the rules of physics (Maxwell's equations) directly.

The Problem:
While the artist is flexible and doesn't need a grid, it has a bad habit. If you ask it to paint a wave that lasts for a long time, it tends to get tired. It starts to "forget" the energy. The waves get smaller and smaller (energy drift), or it might accidentally paint a ripple that moves backward in time (violating causality). It's like an artist who gets so focused on making the picture look good right now that they forget the laws of physics, causing the painting to fall apart over time.

The Solution: A Hybrid Approach
This paper introduces a new way to train this "artist" so it performs as well as the strict "bucket" method (FDTD). The author, Nilufer K. Bulut, uses three clever tricks to fix the artist's bad habits:

1. The "Step-by-Step" Storytelling (Time Marching)

Instead of asking the artist to paint the entire movie of the wave from start to finish in one go, the researchers break the movie into short scenes (time windows).

• The Analogy: Imagine writing a novel. If you try to write the whole book in one sitting, you might forget the plot from Chapter 1 by the time you reach Chapter 20. Instead, you write Chapter 1, finish it, and then use the ending of Chapter 1 as the starting point for Chapter 2.
• The Fix: The AI learns the wave for a tiny slice of time, then uses that result as the "starting line" for the next slice. This prevents the AI from getting confused about cause and effect (causality).

2. The "Seamless Stitch" (Interface Continuity)

When you stitch two pieces of fabric together, you don't want a jagged edge where the pattern jumps.

• The Analogy: When the AI finishes "Scene 1" and starts "Scene 2," there's a risk the wave will jump or glitch at the boundary. The researchers added a special rule (a "loss function") that acts like a seamstress, constantly checking that the wave at the end of one scene matches perfectly with the start of the next.

3. The "Energy Wallet" (Poynting Regularizer)

This is the most important trick. In physics, energy cannot be created or destroyed; it just moves around. But the AI, being a "lazy optimizer," might try to cheat by making the waves disappear to make the math easier.

• The Analogy: Imagine the AI has a wallet where it keeps track of the total energy. The researchers added a "bouncer" (the Poynting regularizer) that checks the wallet at every single point in the room.
• The Twist: The paper discovered that checking the total wallet balance for the whole room (Global) wasn't enough. The AI could cheat by stealing energy from the left side of the room and giving it to the right side, keeping the total balance the same while breaking the laws of physics locally.
• The Fix: Instead of checking the total, the bouncer checks every single person's pocket (Local Poynting). This ensures that energy is conserved everywhere, not just on average. This stopped the "energy drift" completely.

The "Parenthesis Effect" (A Quirky Discovery)

The paper also found something weird and funny. In computer code, sometimes adding or removing a pair of parentheses () changes the result, even if the math looks the same.

• The Analogy: It's like telling a chef, "Add salt, then pepper, then sugar" vs. "Add (salt then pepper), then sugar." The ingredients are the same, but the order of operations in the chef's brain changes slightly.
• The Result: The AI was so sensitive that a tiny change in how the code was written changed how the energy behaved over time. This shows that AI physics solvers are much more sensitive to "how you ask the question" than traditional methods.

The Results

By using these tricks, the AI (PINN) became a master painter.

• Accuracy: It predicted the waves with 99.9% accuracy, matching the strict "bucket" method (FDTD) almost perfectly.
• Energy: It kept the energy wallet balanced with only a 0.024% error, which is incredibly small.
• No Cheating: It did all this without looking at any pre-made answers (labeled data). It learned purely by understanding the rules of physics.

Conclusion

This paper proves that AI can be a powerful tool for simulating electromagnetic waves, but it can't just be "set and forget." You have to build guardrails (like time-marching and local energy checks) to keep it honest. When done right, this "shape-shifting artist" can paint physics just as accurately as the old, rigid "bucket" methods, but with the added benefit of being able to solve tricky, inverse problems that the buckets can't handle.

Technical Summary

1. Problem Statement

While Physics-Informed Neural Networks (PINNs) offer mesh-free advantages and suitability for inverse problems, their application to time-dependent electromagnetic wave propagation (Maxwell's equations) has historically faced significant hurdles compared to established methods like Finite-Difference Time-Domain (FDTD) and Finite Element Method (FEM). Key challenges include:

• Causality Violation: Standard PINNs treat time as a spatial dimension, allowing information to flow backward in time during training, which violates the hyperbolic nature of wave equations.
• Cumulative Energy Drift: In sequential time-window training, small errors in early windows propagate and accumulate, leading to non-physical energy decay or growth.
• Spectral Bias: Neural networks tend to learn low-frequency components faster than high-frequency ones, hindering accuracy in high-frequency regimes.
• Accuracy Gap: Previous PINN implementations often failed to match the field accuracy and energy conservation metrics of classical solvers like FDTD.

The study aims to develop a hybrid methodology that bridges this gap, achieving FDTD-level accuracy and strict energy consistency for 2D Transverse Magnetic (TMz) wave propagation in a Perfect Electric Conductor (PEC) cavity.

2. Methodology

The authors propose a comprehensive hybrid training strategy that integrates theoretical physics constraints directly into the neural network optimization process.

A. Theoretical Framework

• Problem Formulation: The study focuses on the 2D TMz mode of Maxwell's equations in a source-free, lossless PEC cavity. The system is reduced to three coupled PDEs involving $E_z$, $H_x$, and $H_y$.
• Non-dimensionalization: Variables are normalized to ensure numerical stability and balanced gradient magnitudes across different physical quantities (electric vs. magnetic fields).

B. Implementation Strategies

The core innovation lies in the implementation of three specific mechanisms to address the identified challenges:

1. Causality-Aware Time Marching:

   • Time Decomposition: The temporal domain $[0, T_{max}]$ is split into sequential windows (e.g., 20 windows of $\Delta t = 0.1$).
   • Sequential Training: Each window is trained only after the previous one is completed. The final state of window $i$ serves as the initial condition for window $i+1$.
   • Causality-Aware Weighting: To prevent error accumulation, PDE residuals within a window are weighted by an exponential decay function $w_c(\tau) = \exp(-\gamma\tau)$. This forces the optimizer to prioritize accuracy at the beginning of each time window, ensuring a consistent initial regime before errors can propagate.

2. Interface Continuity Loss:

   • To ensure smooth transitions between time windows, a two-stage continuity mechanism is employed.
   • Stage 1: The standard Initial Condition (IC) loss at the start of a window matches the predicted fields against the trained solution of the previous window.
   • Stage 2: A dedicated Interface Continuity Loss is applied on a specific sampling set at the window boundary, minimizing the squared error between the current window's prediction and the previous window's output.

3. Poynting-Based Regularization (Energy Conservation):

   • The authors introduce a Local Poynting Loss ($L_{pyt-local}$) derived from the differential form of Poynting's theorem: $\frac{\partial u}{\partial t} + \nabla \cdot \mathbf{S} = 0$.
   • Unlike global energy constraints (which integrate over the domain and suffer from error cancellation), the local approach enforces energy conservation at every collocation point. This prevents local energy violations from masking each other and accumulating over time.

C. Network Architecture & Training

• Architecture: A deep fully connected network (8 hidden layers, 128 neurons) with skip connections (ResNet-style) and sinusoidal time features.
• Optimization: A two-stage process using Adam (for global structure) followed by L-BFGS (for high-precision fine-tuning).
• Dynamic Weighting: Loss weights are adjusted dynamically during training, prioritizing ICs early on and physical constraints (PDE/Poynting) later.

3. Key Contributions

1. Hybrid Training Framework: A novel combination of time-marching, causality-aware weighting, and interface continuity loss that effectively mitigates the "causality collapse" and error accumulation typical in PINNs.
2. Local Poynting Regularizer: Demonstrates that enforcing energy conservation locally (pointwise) is superior to global integral constraints for preventing cumulative energy drift in sequential training.
3. FDTD-Level Accuracy: Achieves results comparable to FDTD without using any labeled field data for training, relying solely on physics-based residuals.
4. The "Parenthesis Effect": A unique observation that algebraically equivalent code expressions (specifically regarding the grouping of terms in the divergence calculation) can lead to significantly different optimization dynamics and energy behaviors due to floating-point non-associativity and computational graph topology.

4. Results

The model (referred to as pinnA) was evaluated against a high-resolution FDTD reference in several scenarios:

• Field Accuracy:

  • Achieved an average Normalized Root Mean Squared Error (NRMSE) of 0.09% across all field components ($E_z, H_x, H_y$).
  • Average relative L2 error of 1.01%.
  • Performance remained stable across the entire time interval $t \in [0, 2.0]$.

• Energy Conservation:

  • In the lossless PEC cavity, the relative energy mismatch between PINN and FDTD was only 0.024% on average, with a maximum of 0.058%.
  • The model exhibited non-monotonic error behavior, actively correcting energy drops in subsequent windows rather than accumulating drift.

• Scenario Variations:

  • Lossy Medium: The model successfully tracked the expected energy dissipation curve in a conductive medium with a relative error below 0.2%.
  • High Frequency: Even with a dominant frequency increase from ~3.9 to ~8.8 (approx. 17.6 periods), the model maintained an NRMSE of 0.10%, though performance degraded slightly at very high frequencies.
  • Ablation Studies:
    • Removing the Poynting constraint (pinnWP) led to moderate energy drift.
    • Using a Global Poynting constraint (pinnGP) performed worse than having no constraint at all, confirming that global constraints allow error cancellation that degrades local consistency.
    • The Local constraint (pinnA) was essential for stability.

5. Significance

This paper establishes that PINNs can be a viable, competitive alternative to classical numerical solvers (FDTD/FEM) for electromagnetic wave propagation, provided specific hybrid strategies are employed.

• Bridging the Gap: It moves PINNs from a theoretical curiosity to a practical tool capable of matching the accuracy of industry-standard solvers.
• Physical Consistency: By explicitly enforcing causality and local energy conservation, the study ensures that the neural network solutions are not just mathematically close to the PDEs but physically consistent.
• Implementation Sensitivity: The discovery of the "Parenthesis Effect" highlights a critical, often overlooked aspect of PINN development: the specific implementation of mathematical expressions can drastically alter optimization landscapes, a sensitivity absent in classical discretization methods.

In conclusion, the study demonstrates that with a carefully designed hybrid loss function and training pipeline, PINNs can solve complex, time-dependent electromagnetic problems with high fidelity and strict adherence to physical laws.