Usually, traditional machine learning (ML) methods use only labeled samples for learning. However, in practical problems including electromagnetic optimization design, the acquisition cost of labeled samples is relatively high.Obtaining label training samples is the most time-consuming part, so how to use relatively few label samples for training to obtain a high-precision surrogate model is a hot topic. This study proposes a co-training algorithm of semisupervised Gaussian Process (GP) with different kernel functions, based on the differences between these two different GP models. The algorithm is conducted by a small number of labeled samples in combination with unlabeled samples, so as to continuously improve the accuracy of the models. Stop criteria is set in advance to control the number of unlabeled samples introduced, preventing the accuracy of the model reduced by introducing too much unlabeled samples. Furthermore, the proposed algorithm is evaluated by benchmark functions and resonant frequency modeling problems of two different antennas. Results show that the proposed GP model has good fitting effects on the benchmark functions. For the problems of resonant frequency modeling, in the case of the same labeled samples, its predictive ability is better than that of the traditional supervised learning (SL) method.