<p>In this paper, we report a recent experimental study of streak structures in the turbulent separated shear flow by data-driven deep neural network. By applying spanwise-aligned tetrahedron vortex generators upstream of a plane backward-facing step, spanwise-aligned high- and low-speed streaks were generated within the separated shear layer behind the step. The velocity profiles of the shear flow were measured by single-probe hot-wire anemometer in both the streamwise-vertical and the streamwise-spanwise planes in the wind tunnel. Deep neural network models are trained and verified based on the experimental datasets. The input parameter sets include the vortex generator height, spanwise spacing, and the spatial coordinates within the measurement domain, while the output parameter sets are mean and root-mean-square velocities of the shear flow. Mean squared errors between the model-predicted and experimentally measured data are used for quality evaluation of different deep neural network model designs, among which the minimum error of the optimal design descends less than 1%. For other vortex generator parameters, which are not measured in the wind tunnel or used in the training, the model prediction provides reasonable mean velocity contours with streak structures. Thus, we find that the experimental data-driven modeling approach shows reliable robustness for nonlinear fitting of complex datasets as well as considerable generalization for turbulent coherent structures.</p>