[CCoE Notice] MS Thesis Defense Announcement - Yawei Su

[ccoecomm at Central.UH.EDU]

[Dissertation Defense Announcement at the Cullen College of Engineering]
Physics Information Machine Learning For Time Domain Electromagnetic Simulation

Yawei Su

07/27/2022; 2:00PM-4:00PM (CST)
Location: ECE Large Conference Room


Committee Chair:
Dr. Jiefu Chen

Committee Members:
Dr. Xuqing Wu | Dr. David R. Jackson

Abstract

Neural Network has been widely used in all fields of research and has achieved great success. Physics Informed Neural Works (PINNs) is an import and relatively new attempt among deep neural network applications. PINN relays on the universal approximation theorem and aims at solving physics differential equations, providing an alternative way to solve physics problems apart from conventional numerical simulation methods like Finite-difference Time-Domain (FDTD) method or Finite Element Analysis (FEA), turning the solving of equations to an optimization process. In this thesis paper, I follow the previous work of Pan Zhang which tests PINN on several time-domain electromagnetic simulation examples and I concentrate on a simple 1-dimension electromagnetic cavity model with isotropic and homogeneous media and extend the training time range to test the PINN performance on larger time scale problem. We do find PINN shows less probability to converge to an expected physical solution with time scale gets larger and finally fails to give any meaningful result, which is known as the `spectrum bias'. To investigate this failure, we firstly define a parameter `Threshold Period Number' (TPN) and change the PINN parameters to see how these value influence the TPN. Then we change and improve the PINN structure, trying to avoid the failure for larger time scale problem. We also visualize the PINN training process to help us analysis the problem. Although we still can not totally settle the spectrum bias, we find out some preliminary method to practically increase TPN and get better result within certain time scale.

[Engineered For What's Next]
<mailto:riward at central.uh.edu>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://Bug.EGR.UH.EDU/pipermail/engi-dist/attachments/20220719/b276c13a/attachment.html