Loading...
Scalable pipeline for antenna performance prediction based on data-informed machine learnng methods, A
Chen, Yiming
Chen, Yiming
Citations
Altmetric:
Advisor
Editor
Date
Date Issued
2024
Date Submitted
Collections
Research Projects
Organizational Units
Journal Issue
Embargo Expires
Abstract
With the rapid deployment of fifth-generation communication (5G) and the exploration of sixth-generation (6G) communication, antenna engineers face increasing challenges in antenna design and integration. Before the fabrication, antenna engineers must establish specific objective functions with customized constraints based on the design requirements. Then, full-wave simulation tools and search algorithms are involved in optimizing the antenna. During the optimization, the antenna configurations are regarded as the inputs and the performance as the outputs. The trial-and-error iterations are utilized to minimize the cost between the current output and the design requirements. However, the full-wave simulation in each iteration is computationally intensive, and the search algorithms need to optimize their parameters according to the input properties and the customized value ranges; both processes are time-consuming. When there are multiple input variables, antenna engineers always trade off the inputs to get a convergent solution.
The research presented here, based on the machine-learning (ML) point of view, provides a scalable pipeline with different data-informed machine-learning (DIML) workflows. The scalable pipeline here means an efficient objective to accelerate antenna design, in which a series of automated processes are established to reduce the optimization iterations. The pipeline explores the uncovered I/O relationship by the ML models as the probability estimators. The heuristic properties of the well-trained ML models provide reasonable performance predictions based on the random combination of inputs in near real-time matters, even if those combinations are not covered by the full-wave simulations. With the parallel-computing power of the graphic processing unit (GPU), the pipeline takes advantage of hardware acceleration from the beginning of electromagnetic (EM) full-wave simulation, data collection, and post-processing to the final ML validation. By applying the well-trained DIML models, the automatic pipeline provides reasonable prediction as the reference for antenna engineers during the design and integration process. This could reduce the full-wave simulation trials and minimize the iterations before finalizing the antenna configuration for fabrication.
By discussing one of the most important factors of antenna performance, reflection coefficients S11, the pipeline achieves performance prediction of the patch antennas in a wide frequency range. The pipeline is proposed to show its generalization property, which makes it easy to implement on other designs. The fully automated simulation with data collection and the customized ML architecture provide the pipeline powerful scalability in further work with more antenna types and materials, more performance requirements, and wrapping as a pre-trained ML model for other antenna designs.
Associated Publications
Rights
Copyright of the original work is retained by the author.