In this work, a tensor basis neural network was combined with another neural network to predict Reynolds stress in two cases: a flow over a surface-mounted solid block, and over a surface-mounted porous block. Many flow scenarios can be represented by these configurations, including airflow over vegetation, buildings, vehicles and electronic components. These problems contain regions of adverse pressure gradients, recirculation, and stagnation at the front face of the block – all of which are difficult to predict by RANS. To extract data for training, validation and testing, RANS and LES simulations of these two cases at three Reynolds numbers = 1800, 3600 and 7200 based on inlet velocities 1, 2 and 4 m/s and block height = 18 mm were performed using OpenFOAM. After training the machine learning models on data from the Re = 1800 case and validating with the Re = 7200 case, a prediction of the Reynolds stress field was made for the Re = 3600 case. With the LES data taken as ground truth, the mean squared error of predicted τ11 was 90% less compared to the RANS predictions and 60% less for τ12. These improvements demonstrate that we are now able to produce a rapid and accurate inference on Reynolds stress field in the a priori stage for flow problems that are traditionally very challenging to model well with RANS. The wide relevance of the chosen flow cases has already been outlined above and the applicability of this and other augmented RANS approaches to more complex flows shows exciting prospects.