Abstract: In this study, spanwise-aligned Vortex Generators (VGs) were used on the backward-facing step for flow separation control in a low-speed wind tunnel, and the downstream separated shear flow was measured by the single-probe hot-wire anemometer. Based on the experimental dataset, Deep Neural Network (DNN) models were established and trained. The input parameters included the heights and spacings of the vortex generators and the spatial coordinates within the separated shear layer, while the output parameters were time-averaged velocities and the root-mean-square of the turbulent fluctuations. The experimental datasets were used in the iterating and training process of the models. In the comparison of the model-predicted and measured results, the mean squared errors between the modeled mean velocities and fluctuation distributions and the measured data at various flow positions were all below 10⁻³, and thus the data-driven deep neural network models were validated with the ability of nonlinear modeling of the separated shear flow, characterization of flow features and prediction of experimental results. Furthermore, the DNN model managed to enrich the complex flow field information by nonlinear fitting of the discrete measurement points, which leads to optimization of vortex generator designs.