Fast machine-learning-enabled size reduction of microwave components using response features.

Area of Science:

• Electrical Engineering
• Computational Electromagnetics

Background:

• Compact size is crucial in modern microwave design.
• Miniaturization requires precise parameter tuning and global optimization.
• Electromagnetic (EM) simulation for evaluating compact structures is computationally expensive.

Purpose of the Study:

• To introduce an innovative and cost-effective method for explicit global miniaturization of microwave components.
• To address the challenges of expensive constraints and system evaluation in size reduction problems.

Main Methods:

• Leveraging response feature technology with characteristic points from EM-analyzed responses.
• Employing an implicit constraint handling approach to transform the problem into an unconstrained one.
• Utilizing a machine-learning framework with kriging-based surrogates and predicted improvement as the infill criterion.

Main Results:

• Demonstrated effectiveness on two miniaturized coupler designs.
• Outperformed conventional gradient-based, population-based, and other machine learning benchmark routines.
• Achieved optimization costs typically between 100 to 150 EM circuit analyses.

Conclusions:

• The proposed framework offers a reliable and computationally efficient solution for microwave component miniaturization.
• The method simplifies the setup process for complex optimization problems.
• Enables significant size reduction while meeting stringent performance requirements.