Aiming at reducing the large computation cost of traditional EM-driven antenna design methods, surrogate models based on back propagation neural networks (BPNN) are studied. In order to solve the problem of easily falling into local optimum in BPNN, a PSO-BPNN surrogate model is developed by improving initial structural parameters of neural networks and applied to fast multi-objective optimization design of multi-parameter antenna structures. Design results show that the proposed PSO-BPNN surrogate model can be integrating into multi-objective evolutionary algorithms for dealing with complex antenna designs with high-dimensional parameter space.