Machine-learning-based global optimization of microwave passives with variable-fidelity EM models and response...

Area of Science:

• Electrical Engineering
• Computational Electromagnetics
• Machine Learning Applications

Background:

• Optimizing microwave passive components is crucial for complex modern circuits.
• Full-wave electromagnetic (EM) simulations are computationally expensive for global optimization.
• Existing methods struggle with nonlinear characteristics and high computational demands.

Purpose of the Study:

• To develop a cost-effective global parameter tuning technique for microwave passive components.
• To address the computational burden of traditional EM-driven optimization.
• To improve the efficiency of machine learning-based design optimization.

Main Methods:

• Leveraging variable-fidelity electromagnetic (EM) simulations.
• Employing response feature technology within a kriging-based machine learning framework.
• Utilizing a co-kriging surrogate model and a particle swarm optimizer.

Main Results:

• The proposed framework achieves competitive design quality and computational cost.
• It requires significantly fewer high-fidelity EM analyses compared to benchmark methods (typically sixty).
• Demonstrates efficacy against nature-inspired algorithms, gradient search, and direct ML techniques.

Conclusions:

• The innovative technique offers an efficient solution for microwave passive component parameter tuning.
• Variable-fidelity simulations and response features reduce computational expense.
• The framework provides a practical approach for complex circuit design optimization.