To speed up the optimization of antenna modeling, this paper proposes a novel deep multi-layer perceptron (DMLP) network based on deep learning network architecture for optimizing ultra-wideband antenna. The DMLP network uses a step-down, connected-layer deep network, and the Adam optimizer automatically updates the learning rate. Dropout technology is used to remove random neurons in the hidden layer, preventing overfitting due to the deep network layers. This paper uses the DMLP network to model the geometric parameters of the ultra-wideband stepped microstrip monopole antenna, extracts features from the eight geometric parameters of the antenna, and predicts the S11 value of the antenna. The experimental results show that compared with traditional multilayer perceptron and radial-basis-function neural networks, the average prediction error of S11 is reduced by 118.32% and 123.76%, respectively, and it has a higher prediction accuracy. In addition, the fitting speed is improved. The feasibility of this network is verified through experiments.