近年來，深度學習技術蓬勃發展，為電磁模擬領域帶來新的研究契機。本論文探討物理訊息神經網路（Physics-Informed Neural Networks, PINNs）於電磁波模擬問題中的應用，選定具散射體之平行波導結構為研究對象，建構以 Maxwell's 方程為基礎的 PINNs 模型，模擬不同結構下的電場分佈，並與傳統有限元素法（Finite Element Method, FEM）進行比較分析。
本研究分別探討兩種幾何設定：其一為單散射體結構，探討網路超參數（如層數、神經元數量、激活函數...等）對模型學習場型精度的影響；其二雙散射體結構，分析不同夾縫尺寸對場傳播與干涉現象的影響，進一步驗證 PINN 模型於多重散射問題中的學習能力與準確性。
結果顯示，透過合理設計網路架構與超參數設定，可提升模型在高梯度變化區域之擬合能力，並有效重建主模態與局部增強場。此外，在較複雜的雙散射體情境中，PINN 模型雖仍面臨局部場不易收斂之挑戰，但整體場分佈趨勢與 FEM 結果一致，展現深度學習結合物理模型於電磁模擬應用上的潛力與可行性。
本研究驗證了 PINN 架構在具邊界條件與結構複雜性的波導模擬中的適用性，亦為未來拓展至三維模擬、多物理耦合與逆問題解法奠定基礎。

In recent years, the rapid advancement of deep learning technologies has opened new research avenues in electromagnetic simulation. This thesis investigates the application of Physics-Informed Neural Networks (PINNs) in modeling electromagnetic wave phenomena, focusing on parallel waveguide structures containing perfect electric conductor (PEC) scatterers. A PINN framework based on Maxwell’s equations in the frequency domain is constructed to predict the electric field distributions under different geometries, and the results are compared with those obtained using the traditional Finite Element Method (FEM).
Two types of geometrical configurations are considered in this study. The first involves a single centrally placed PEC scatterer forming an upper and lower channel, where the effects of neural network hyperparameters—such as the number of layers, neurons per layer, and activation functions—on prediction accuracy are systematically analyzed. The second configuration consists of two symmetric PEC scatterers forming a narrow slit channel, in which various slit dimensions are examined to evaluate the model’s capacity to capture wave propagation and interference effects in multi-scattering environments.
The results demonstrate that with appropriate design of the network architecture and loss function weighting, the PINN model can effectively reconstruct the main field modes and local field enhancements, especially in regions with steep field gradients. Although challenges such as local non-convergence remain in the double-scatterer configuration, the overall field distribution predicted by PINNs is consistent with FEM results, highlighting the potential and feasibility of combining deep learning with physical modeling in electromagnetic simulations.
This work verifies the applicability of the PINN framework for simulating guided-wave problems with complex geometries and boundary conditions, laying a foundation for future extensions to three-dimensional modeling, multiphysics coupling, and inverse design problems.

[1] M. Rasht-Behesht, C. Huber, K. Shukla, and G. E. Karniadakis, “Physics-informed neural networks (pinns) for wave propagation and full waveform inversions,” Journal of Geophysical Research: Solid Earth, vol. 127, no. 5, 2022.
[2] M. Raissi, P. Perdikaris, and G. E. Karniadakis, “Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations,” Journal of Computational physics, vol. 378, 2019.
[3] G. Pang, L. Lu, and G. E. Karniadakis, “Fpinns: Fractional physics-informed neural networks,” SIAM Journal on Scientific Computing, vol. 41, no. 4, 2019.
[4] G. E. Karniadakis, I. G. Kevrekidis, L. Lu, P. Perdikaris, S. Wang, and L. Yang, “Physics-informed machine learning,” Nature Reviews Physics, vol. 3, no. 6, 2021.
[5] Z. Mao, A. D. Jagtap, and G. E. Karniadakis, “Physics-informed neural networks for high-speed flows,” Computer Methods in Applied Mechanics and Engineering, vol. 360, 2020.
[6] C. A. Balanis, Advanced engineering electromagnetics. John Wiley & Sons, 2012.
[7] J. D. Jackson and R. F. Fox, Classical electrodynamics, 1999.
[8] D. J. Griffiths, Introduction to electrodynamics. Cambridge University Press, 2023.
[9] J. D. Jackson and R. F. Fox, Classical electrodynamics, 1999.
[10] M. N. Sadiku and S. Nelatury, Elements of electromagnetics. Oxford university press New York, 2001, vol. 428.
[11] A. Taflove, S. C. Hagness, and M. Piket-May, “Computational electromagnetics: The finite-difference time-domain method,” The Electrical Engineering Handbook, vol. 3, 2005.
[12] K. S. Kunz and R. J. Luebbers, The finite difference time domain method for electromagnetics. CRC press, 1993.
[13] K. S. Yee, “Numerical solution of initial boundary value problems involving maxwell’s equation,” IEEE Trans, an Antenna and propagation, vol. 14, pp. 302–307, 1966.
[14] S. D. Gedney, “An anisotropic perfectly matched layer-absorbing medium for the truncation of fdtd lattices,” IEEE transactions on Antennas and Propagation, vol. 44, no. 12, pp. 1630–1639, 1996.
[15] J.-P. Berenger, “A perfectly matched layer for the absorption of electromagnetic waves,” Journal of computational physics, vol. 114, no. 2, pp. 185–200, 1994.
[16] O. C. Zienkiewicz and R. L. Taylor, The finite element method set. Elsevier, 2005.
[17] ¨O. ¨Ozg¨un and M. Kuzuo˘glu, MATLAB-based finite element programming in electromagnetic modeling. CRC Press, 2018.
[18] J.-M. Jin, The finite element method in electromagnetics. John Wiley & Sons, 2015.
[19] P. P. Silvester and R. L. Ferrari, Finite elements for electrical engineers. Cambridge university press, 1996.
[20] G. E. Karniadakis, I. G. Kevrekidis, L. Lu, P. Perdikaris, S. Wang, and L. Yang, “Physics-informed machine learning,” Nature Reviews Physics, vol. 3, no. 6, pp. 422–440, 2021.
[21] S. Schoder and F. Kraxberger, “Feasibility study on solving the helmholtz equation in 3d with pinns,” arXiv preprint arXiv:2403.06623, 2024.
[22] S. Nair, T. F. Walsh, G. Pickrell, and F. Semperlotti, “Multiple scattering simulation via physics-informed neural networks,” Engineering with Computers, pp. 1–20, 2024.