Inverse Design Method for Horn Antennas Based on Knowledge-Embedded Physics-Informed Neural Networks

This letter aims to overcome the challenges faced by physics-informed neural networks (PINNs) in metal structure design. Taking the design of horn antennas as an example, we propose knowledge-embedded PINNs inverse design framework. Normalized Maxwell's equations are employed to address converg...

Full description

Saved in:
Bibliographic Details
Published in:IEEE antennas and wireless propagation letters 2024, Vol.23 (6), p.1665-1669
Main Authors: Liu, Jin-Pin, Wang, Bing-Zhong, Chen, Chuan-Sheng, Wang, Ren
Format: Article
Language:English
Subjects:
Apertures
Biological neural networks
Boundary conditions
Electric fields
Horn antennas
Inverse design
inverse problems
Mathematical models
Maxwell equations
Maxwell's equations
Metals
Neural networks
physics-informed neural networks (PINNs)
topology optimization
Training
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:This letter aims to overcome the challenges faced by physics-informed neural networks (PINNs) in metal structure design. Taking the design of horn antennas as an example, we propose knowledge-embedded PINNs inverse design framework. Normalized Maxwell's equations are employed to address convergence issues caused by differences in magnitudes. Unlike traditional PINNs, we built field component neural networks, inherently satisfying the boundary conditions of the metal structure. At the same time, the boundary conditions of the port and aperture fields are embedded within the neural network using a hard constraint boundary approach. The embedded knowledge further simplifies the loss function. The numerical experiments showcase the inverse design results of the H-plane horn antenna profile under different target aperture field requirements. The results demonstrate that the inverse-designed antennas effectively achieve the super-gain and beam deflection, respectively, validating the feasibility and practical value of the proposed method.
ISSN:1536-1225
1548-5757
DOI:10.1109/LAWP.2024.3365690