In this work, we introduce a bi-stage Bayesian framework for multi-fidelity optimization of a patch antenna. The first stage involves exploration of the search space using a set of independent trust region (TR) based optimization runs executed on separate Gaussian process (GP) models, each initialized with a random set of samples [1]. The evaluation of data-points is performed using a numerically cheap low-fidelity model. In the course of the exploration phase, the meta-algorithm governs the individual TR instances towards the most promising regions of the search space. In the second stage, the best identified region is exploited using a multi-fidelity method that blends the information from the low- and high-fidelity (HF) simulations of the antenna into another GP [2]. Data fusion facilitates identification of a high-quality design solutions using only a handful of expensive model simulations. The initial results demonstrate a rapid convergence to the optimal solutions, characterized by objective values below –10 dB. Moreover, in the exploitation phase, the objective function exhibits a noticeable decrease while increasing the number of HF iterations. The main advantage of the proposed method is that, contrary to conventional TR algorithms, it does not require engineering insight, or a priori information on the problem at hand (here, a good starting point for antenna optimization) to identify the promising solutions [3]. Secondly, the algorithm seamlessly integrates exploitation and exploration strategies while being computationally cheaper compared to competitive unassuming approaches such as the state-of-the-art population-based metaheuristic algorithms.