Although the technological advances that led to the development of fast computers, the direct numerical simulation of turbulent flows is still prohibitively expensive to most engineering and even some research applications. The CFD simulations used worldwide are, therefore, based on averaged quantities and heavily dependent on mathematical turbulence models. Despite widely used, such models fail to proper predict the averaged flow in many practical situations, such as the simple flow in a square duct. With the re-blossoming of machine learning methods in the past years, much attention is being given to the use of such techniques as a replacement to the traditional turbulence models. The present work evaluated the use of Neural Networks as an alternative to enhance the simulation of turbulent flows. To this end, the Stereoscopic-PIV technique was used to obtain well-converged flow statistics and velocity fields for the flow in a square duct for 10 values of Reynolds number. A total of 10 methodologies were evaluated in a data-driven approach to understand what quantities should be predicted by a Machine Learning technique that would result in enhanced simulations. From the selected methodologies, accurate results could be obtained with a Neural Network trained from the experimental data to predict the nonlinear part of the Reynolds Stress Tensor and the turbulent eddy viscosity. The turbulent simulations assisted by the Neural Network returned velocity fields with less than 4 percent in error, in comparison with those previously measured.