Antennas Modelling based on Consensus Gaussian Process

When optimizing the antennas, the traditional method is using the swarm-based global optimal algorithms combining full-wave electromagnetic simulation software. The biggest shortcoming of the method is low efficiency because it needs many times calling the simulation software which is very time-consuming. In this study, we adopt machine learning methods to replace EM simulation software, fasting the optimal process of the antennas, in which the most important thing is its accuracy and generalization ability. We propose an ML surrogate model based on the consensus concept exploiting Gaussian process. Specifically, we model the optimized antenna independently several times by GPs, considering that GP is suitable for the small sample antenna issue. Then, we obtain the final surrogate model by either consistency or aggregation based on the trained GPs, where a consistency algorithm is used when the difference between the GPs is small whereas an aggregation algorithm is used when the difference between the GPs is big, in order to guarantee the good reliability and excellent accuracy of the results. The experiments of an ultra-wideband antenna and a multiple-input and multiple-output antenna demonstrate the proposed method with great generalization ability, where the evaluation metrics are excellent.