Many full-wave electromagnetic (EM) simulations are needed to design an antenna meeting certain requirements, which involves a considerable computational burden. A multibranch machine learning-assisted optimization (MB-MLAO) method is proposed to dramatically reduce the computational complexity involved in this task. This method is then applied to antenna design and worst-case performance (WCP) searching under a practical manufacturing tolerance. In the conventional Gaussian process regression (GPR)-based MLAO method, a lower confidence bound (LCB) prescreening strategy with an empirical LCB constant is used to weigh the predicted value and predicted uncertainty. Using a variable-fidelity machine learning method, an adaptive LCB variable, and a retraining and repredicting method, the proposed MB-MLAO method can strike a delicate balance between exploitation and exploration in searching. Moreover, variable-fidelity data from full-wave EM simulations are used in the deep GPR machine learning method to further reduce the computational burden. Finally, two test functions and four types of antennas are selected as examples to illustrate the superiority of the proposed MB-MLAO method.