A black box model is a numerical simulation that is used in optimization. It is computationally expensive, so it is convenient to replace it with surrogate models obtained by simulating only a few points and then approximating the original black box. Here, a recent approach, using Symbolic Regression via Genetic Programming, is compared experimentally to neural network based surrogate models, using test functions and electromagnetic models. The accuracy of the model obtained by Symbolic Regression is proved to be good, and the interpretability of the function obtained is useful in reducing the optimization's search space. © 2011 Authors.