Abstract
Automated model generation (AMG) is an automated artificial neural network (ANN) modeling algorithm, which integrates all the subtasks (including adaptive sampling/data generation, model structure adaptation, training, and testing) in neural model development into one unified framework. In existing AMG, most of the time is spent on data sampling and model structure adaptation due to the iterative neural network training and the sequential computation mechanism. In this paper, we propose an advanced AMG algorithm using parallel computation and interpolation approaches to speed up the neural modeling of microwave devices. Efficient interpolation approaches are incorporated to avoid repetitive training of the intermediate neural networks during adaptive sampling process in AMG. Parallel computation formulation based on a multi-processor environment is proposed to further save time during interpolation calculation, data generation, and model structure adaptation process. Examples of automated modeling of two microwave filters are presented to show the advantage of this paper.
Introduction: The Bottleneck in Microwave Design
In modern microwave engineering, Electromagnetic (EM) simulators are indispensable for predicting device behavior with high precision. However, these simulators are computationally "expensive" and slow, which hinders the iterative design and optimization process. Artificial Neural Networks (ANNs) offer a compelling solution: once trained on EM data, they can provide near-instantaneous results, acting as high-speed surrogate models.
Despite their benefits, building these ANN models is itself a significant bottleneck. The standard approach, known as Automated Model Generation (AMG), employs "adaptive sampling"—an iterative process that adds data points specifically in regions where the model is least accurate. In traditional AMG, the system must retrain the entire neural network every time new data is added just to determine where to sample next. This sequential computation mechanism is incredibly time-consuming, especially as the complexity of the microwave device increases.
Methodology: A Parallelized AMG Framework
To address these inefficiencies, this paper proposes an advanced AMG algorithm that integrates parallel computation and localized interpolation. The implementation, developed in C using OpenMP for multi-processor management, targets Intel Xeon environments to maximize throughput. The framework is built on three key parallelized pillars:
1. Parallel Interpolation Approach
The core innovation of this research is the realization that a fully trained, optimized ANN is not required to decide where to collect more data. Instead, the algorithm uses simple local interpolation functions to estimate error. Because these local models are independent, they can be calculated across multiple processors simultaneously. This avoids the repetitive and heavy training of intermediate neural networks during the adaptive sampling process.
2. Parallel Data Generation
The system is designed to drive multiple EM simulators (such as CST Microwave Studio) in parallel. By utilizing a multi-processor environment, the algorithm can generate multiple data points in the time it previously took to generate one. For instance, a four-processor setup can theoretically quadruple the data generation speed.
3. Parallel ANN Training
Once the necessary data is collected, the final training of the neural network is distributed. The workload for calculating "feedforward" errors and "backpropagation" derivatives is split across available CPU cores, ensuring that the final model generation is as fast as the data gathering phase.
Adaptive Granularity
As illustrated in the adaptive sampling process, the algorithm automatically divides the input space into subregions ($2^n$ subregions). It focuses computational power on nonlinear, complex regions of the device's behavior while maintaining a coarser grid in simpler, linear regions. This adaptive granularity ensures that resources are spent where they are most needed to achieve high accuracy.
Experimental Evaluation and Results
The proposed parallel AMG was rigorously tested against standard sequential AMG and a modified version using interpolation without parallelism. The study utilized two practical microwave design cases:
- Two-section low-pass elliptic microstrip filter: A 6-parameter model where 1,394 training samples were generated across 101 frequency points.
- Bandpass High-Temperature Superconductor (HTS) microstrip filter: A 6-parameter model also utilizing 1,394 samples.
The results highlight the dramatic efficiency gains of the proposed method:
- Significant Speedup: For the low-pass filter, the total CPU time dropped from 22.69 hours (standard AMG) to just 5.90 hours (proposed method).
- High Parallel Efficiency: The efficiency for data generation and ANN training remained consistently high (above 80%), demonstrating that the algorithm scales effectively with additional CPU cores.
- Uncompromised Accuracy: Despite the massive reduction in wall-clock time, the final models maintained the user-desired 1% testing error. This proves that the shortcuts taken during the sampling phase—using interpolation instead of full training—did not compromise the integrity of the final model.
Discussion and Future Directions
By moving from a sequential "Sample -> Train -> Test -> Repeat" loop to a parallelized "Parallel Sample -> Parallel Interpolate -> Parallel Train" flow, the authors have transformed a linear time problem into one that scales with hardware.
However, some limitations remain. The number of subregions grows exponentially ($2^n$) with the number of physical parameters, which could become a bottleneck for very high-dimensional problems (e.g., 20+ parameters). Future work could explore adapting this framework for GPUs, which are inherently better suited for the matrix mathematics required for ANNs. Additionally, making this framework "plug-and-play" for a wider variety of commercial EM simulators remains a practical challenge for industry-wide adoption.
In conclusion, the proposed parallel AMG algorithm offers a robust and efficient path for the automated development of high-speed neural models, significantly reducing the time-to-market for complex microwave components.
Summary updated