This thesis aims to explore the application of machine learning techniques to improve turbulence modeling and numerical simulations in fluid mechanics. More specifically, we are interested in the application of artificial neural networks (ANNs) for large eddy simulation. The latter is a modeling approach that focuses on the direct resolution of large turbulent structures, while modeling small scales by a subgrid-scale model. It requires a certain ratio of total kinetic energy to be resolved. However, this ratio may be difficult to achieve for industrial simulations due to the high computational cost, leading to under-resolved simulations. We aim to improve the latter by focusing work along two main axes: 1) Using ANNs to build generic sub-mesh models that outperform analytical models and compensate for coarse spatial discretization; 2) Training ANNs to learn wall models. One of the main challenges is the ability of the new models to generalize correctly in configurations different from those used during training. Thus, taking into account the different sources and quantification of uncertainties plays a vital role in improving the reliability and robustness of machine-learned models.