In recent years, machine learning techniques (MLTs) have been applied to accelerate the analysis and design of electromagnetic devices. Algorithms such as artificial neural networks or support vector machines for regression (SVRs) have been proposed for the design of large reflectarrays for space applications at a single frequency. However, multifrequency optimization of such large antennas has not been tackled with MLTs. In this letter, for the first time, we propose a technique based on the use of SVR analysis to obtain the reflection coefficients to accelerate the design of a very large shaped-beam reflectarray for direct broadcast satellite in a 15% bandwidth. An in-house method of moments based on local periodicity is employed to generate samples to train the SVRs for each considered frequency. Then, the surrogate model is used for a design at central frequency, which is used as starting point for a wideband design procedure that is accelerated more than an order of magnitude without a significant loss of accuracy. It is shown that, by the virtue of the proposed methodology, the minimum copolar gain in the coverage zone is improved more than 10 dB at the upper frequency while maintaining a computationally efficient design procedure.
- Direct broadcast satellite (DBS)
- Generalized intersection approach (IA)
- Machine learning
- Support vector regression (SVR)
- Wideband reflectarray antenna
ASJC Scopus subject areas
- Electrical and Electronic Engineering