Physics-Guided Neural Surrogate Model with Particle Swarm-Based Multi-Objective Optimization for Quasi-Coaxial TSV Interconnect Design | JoVE Visualize

Area of Science:

Electrical Engineering and Computer Science
Electromagnetics and Computational Electromagnetics
Microsystems and RF Engineering

Background:

Accurate electromagnetic (EM) modeling is crucial for high-frequency signal integrity in reconfigurable radio frequency (RF) microsystems.
Conventional neural network surrogate models often neglect physical constraints like causality and passivity, limiting their real-world applicability.
Existing methods struggle to balance numerical accuracy with essential physical principles for reliable RF microsystem performance prediction.

Purpose of the Study:

To develop a physics-constrained Neuro-Transfer surrogate model for accurate S-parameter prediction in RF microsystems.
To enforce causality and passivity constraints within the surrogate model using dedicated regularization terms.
To optimize the performance of a quasi-coaxial TSV composite structure using a particle swarm optimization (PSO) framework.

Main Methods:

A Neuro-Transfer surrogate model with a broadband output architecture was proposed to predict S-parameters from 1-50 GHz.
Causality and passivity were enforced via regularization terms during the model's training phase.
A multi-objective PSO framework, incorporating fixed-weight normalization and a linearly decreasing inertia weight, optimized S11, S21, and S22 parameters.

Main Results:

The physics-constrained model achieved direct S-parameter prediction over the 1-50 GHz bandwidth.
Optimized structural parameters for the TSV composite structure met target values for S11 (-25 dB), S21 (-0.54 dB), and S22 (-24 dB).
Prediction-to-simulation deviations were below 1 dB, with an average prediction error of 2.11% on the test set.

Conclusions:

The proposed physics-constrained Neuro-Transfer surrogate model effectively predicts S-parameters while adhering to critical physical constraints.
The PSO-based optimization framework successfully improved the performance of the quasi-coaxial TSV composite structure.
This approach enhances the accuracy and reliability of EM modeling for RF microsystem design and development.