Investigating KAN-Based Physics-Informed Neural Networks for EMI/EMC Simulations

Kun Qian, Mohamed KheirarXiv 2024 · DOI · arXiv

Abstract

Traditional electromagnetic (EM) simulations for Electromagnetic Interference (EMI) and Compatibility (EMC) are computationally intensive, often requiring significant time and high-end hardware. This study investigates the application of Kolmogorov-Arnold Networks (KANs) within a Physics-Informed Neural Network (PINN) framework to solve electromagnetic partial differential equations (PDEs). By replacing standard Multi-Layer Perceptrons (MLPs) with KANs—which utilize learnable activation functions on edges—the research demonstrates that comparable accuracy can be achieved with significantly smaller network architectures. This approach offers a path toward "green" EMI simulations by reducing computational overhead and energy consumption without sacrificing physical fidelity.

Introduction

Ensuring Electromagnetic Compatibility (EMC) is a critical yet resource-heavy phase in electronic design. Current "full-wave" numerical solvers for Maxwell’s equations require substantial time, high-end hardware, and significant energy. The iterative nature of design further compounds these costs, as every minor change necessitates a new simulation. This paper explores an AI-driven alternative: Physics-Informed Neural Networks (PINNs) optimized with Kolmogorov-Arnold Networks (KANs) to provide faster, more efficient EM simulations.

Methodology

The proposed framework leverages PINNs, which integrate physical laws directly into the neural network's loss function. This ensures that the model's output adheres to the underlying physics, such as the Laplace equation, rather than relying solely on data fitting.

The primary innovation is the substitution of standard Multi-Layer Perceptrons (MLPs) with Kolmogorov-Arnold Networks (KANs). While MLPs use fixed activation functions on neurons and trainable weights on connections, KANs employ learnable activation functions (typically B-splines) on the connections themselves. This allows the network to capture complex physical behavior more compactly.

Architecture of the KAN-based PINN solution network integrated with a physics network.

Experimental Setup

The effectiveness of the KAN-based PINN was evaluated by solving the 2D Laplace equation ($-\nabla^2 u = 0$) for an electrostatic box. The domain is defined by a unit square with specific Dirichlet boundary conditions: $V = 1V$ at the top edge ($y=1$) and $V = 0V$ on the other three edges.

The 2D electrostatic box problem setup with defined boundary conditions.

The training set included 50 points sampled uniformly along each boundary and 2,500 points sampled randomly within the interior of the domain. Validation was performed against a "gold standard" solution generated by the commercial solver CST Studio Suite on a $101 \times 101$ uniform grid.

Results and Discussion

The KAN-based PINN demonstrated high accuracy, matching the commercial solver's results with an absolute error below 0.1. Notably, the KAN architecture achieved this performance with only 5 neurons per hidden layer, whereas the MLP required 32 neurons per layer to reach similar accuracy.

This reduction in model size highlights several advantages for "green" engineering:

  • Computational Efficiency: Smaller networks require less memory and fewer operations during both training and inference.
  • Sustainability: Reduced training time and computational load directly lower the electricity and cooling costs associated with simulation workflows.
  • Interpretability: KANs can potentially yield symbolic representations of the learned physical functions, making them less of a "black box" than traditional deep MLPs.

Conclusion

This feasibility study confirms that KAN-based PINNs are a viable and efficient tool for solving electrostatic problems in EMI/EMC contexts. While future work must address more complex 3D geometries and full-wave Maxwell equations, the results suggest that KANs can significantly augment traditional simulation tools, especially during early-stage design iterations where speed and efficiency are paramount.

Summary updated

BibTeX
@misc{qian2024investigating,
  author = {Kun Qian and Mohamed Kheir},
  title = {{Investigating KAN-Based Physics-Informed Neural Networks for EMI/EMC Simulations}},
  year = {2024},
  eprint = {2405.11383},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2405.11383},
}
Figure