Abstract
Traditional computational electromagnetics (CEM) relies on the Method of Moments (MoM), which is constrained by rigid, hand-crafted basis functions. This rigidity necessitates high-resolution meshes for complex geometries, leading to significant computational overhead. We present PhyKAN, a physics-informed neural network framework that replaces fixed basis functions with learnable, adaptive ones. By utilizing Kolmogorov-Arnold Networks (KAN) and embedding the dyadic Green’s function as a physical prior, PhyKAN achieves high-accuracy surface current modeling. Experimental results demonstrate that PhyKAN significantly outperforms traditional MoM on coarse meshes, maintaining physical rigor through a self-supervised Electric Field Integral Equation (EFIE) loss.
Introduction
In the field of CEM, solving how electromagnetic waves scatter off objects is typically handled by the Method of Moments. While effective, MoM relies on predefined basis functions (such as Rao-Wilton-Glisson functions) that are fixed to the mesh geometry. To increase accuracy for complex shapes or varying physics, practitioners must refine the mesh, which exponentially increases computational costs.
PhyKAN addresses this "rigidity" problem by making the basis functions themselves learnable and adaptive. Unlike standard deep learning approaches that treat simulations as black boxes, PhyKAN incorporates the underlying physics of the EFIE to ensure the predicted surface currents satisfy the boundary conditions of perfect conductors.
Methodology
The PhyKAN architecture consists of a dual-branch system designed to mirror the mathematical structure of the EFIE:
- Local Branch (Learnable Basis): This branch employs Kolmogorov-Arnold Networks (KAN). Unlike traditional Multi-Layer Perceptrons (MLPs) with fixed node activations, KANs use learnable univariate functions on the edges. This allows the network to adaptively "shape" basis functions based on local geometric features like edge length and face normals.
- Global Branch (Green’s Function Prior): To account for long-range interactions, the dyadic Green’s function is embedded directly into the network. The first layer weights are initialized with the Green’s function and then projected into a higher-dimensional latent space, ensuring the model respects Maxwell’s equations from initialization.
- Physics-Informed Loss: The framework is self-supervised, utilizing a loss function that minimizes the EFIE residual. This ensures that the predicted surface currents are physically consistent without requiring labeled ground-truth data.
Experimental Evaluation
PhyKAN was evaluated on several canonical Perfectly Electrically Conducting (PEC) geometries, including cubes, spheres, cones, and complex benchmark objects like the SLICY model. Simulations were conducted across frequencies ranging from 1 GHz to 5 GHz, with results compared against high-fidelity MoM solutions.
Key Findings
- Superiority on Coarse Meshes: A primary advantage of PhyKAN is its performance on low-resolution meshes. At a mesh size of $\lambda$, PhyKAN reduced the Mean Squared Error (MSE) by approximately 45% compared to traditional MoM. The learnable basis functions can represent complex current variations that rigid functions miss on coarse grids.
- High Fidelity and Robustness: The model achieved reconstruction relative errors below 3% across all tested geometries. Furthermore, it accurately preserved Radar Cross Section (RCS) patterns, essential for antenna and stealth design, despite not being explicitly trained on RCS data.
- Impact of Physical Priors: Ablation studies confirmed that initializing the global branch with the Green’s function is vital; removing this prior caused error rates to more than double.
Conclusion
PhyKAN represents a significant step forward in CEM by bridging the gap between classical physical rigor and the flexibility of neural networks. By learning the basis functions, the framework allows for accurate electromagnetic modeling on significantly coarser meshes than previously possible. Future work will focus on scaling the architecture to massive, multi-platform simulations and extending its capabilities to broadband performance.
Summary updated