Abstract
This paper presents a novel Automated Model Generation (AMG) algorithm for developing neural network-based multiphysics parametric models of microwave components. By integrating electromagnetic (EM)-centric multiphysics data generation, parallel computing, and automated neural network structure adaptation into a unified framework, the proposed technique significantly reduces the human effort and modeling time required by traditional manual methods. Validation on complex microwave filters demonstrates a three-fold speedup in model development while maintaining high predictive accuracy.
Introduction
In modern high-performance microwave design, analyzing electromagnetics (EM) in isolation is no longer sufficient. Engineers must account for "multiphysics" effects—how heat dissipation or structural deformation changes a component's electrical behavior. While Artificial Neural Networks (ANNs) are excellent at creating fast, "surrogate" models of these complex interactions, building them is traditionally a slow, manual process. It requires human experts to manually run hundreds of simulations and use trial-and-error to determine the optimal network architecture. This paper introduces an automated pipeline to remove this "human bottleneck."
The Automated Model Generation (AMG) Framework
The proposed AMG algorithm automates the entire modeling lifecycle, from data collection to final optimization. The process is divided into two primary phases: Parallel Multiphysics Data Generation and Automated Multiphysics Model Structure Adaptation.
1. Automated Parallel Data Generation
The framework includes built-in drivers for multiphysics simulation tools (such as COMSOL). It automatically partitions the required data samples and distributes them across multiple computers. This parallelization significantly reduces the time spent in the most computationally expensive phase of the project, achieving parallel efficiencies between 76% and 81%.
2. Automated Structure Adaptation
Instead of manual guessing, the algorithm uses a stage-wise approach to find the optimal number of hidden neurons. It monitors training and testing errors to detect "under-learning" (adding neurons) or "over-learning" (reducing neurons). The goal is to find the most compact model that meets a user-defined accuracy threshold, ensuring the final model is both accurate and fast to evaluate.
Neural Model Architecture
The resulting neural network acts as a parametric model that maps various input domains to the component's multiphysics response. These inputs include geometric parameters (e.g., length, width), non-geometric parameters (e.g., material properties), and operating conditions like frequency or control voltages.
Experimental Validation and Results
The researchers evaluated the framework using two complex designs: a Tunable Evanescent-Mode Cavity Filter and a Tunable Four-Pole Waveguide Filter.
Key findings include:
- Significant Speedup: For the waveguide filter, total CPU time was reduced from 42.95 hours using manual methods to just 13.96 hours using the automated framework.
- Robust Convergence: Regardless of the initial guess for the number of neurons, the algorithm consistently converged to the same optimal model size.
- High Accuracy: Both models achieved testing errors between 1.5% and 1.95%, meeting the stringent precision requirements of microwave engineering.
Conclusion
The proposed automated technique transforms neural-based multiphysics modeling from a manual, expert-driven task into an efficient computational process. By combining parallel simulation driving with adaptive architecture optimization, the framework provides a scalable solution for high-level multiphysics design in microwave engineering.
Summary updated