The spread of machine learning (ML) techniques in combination with the availability of high-quality experimental and numerical data boosted in recent years numerous applications in fluid mechanics. Among those, examples of closure models for turbulent flows or data-assimilation based on neural networks (NN) are already numerous. However, it is well known that these techniques are prone to over-fit and necessitate an exceedingly large amount of data, unless enforcing physical constraints. We address those limitations by applying graph neural networks (GNN). This architecture is characterized by a net of nodes that can be easily adapted to unstructured meshes. Moreover, it is known GNN can show remarkable generalization capabilities as compared to standard network models. Here, we demonstrate the use of GNN by interfacing them with a finite elements (FEM) solver for the supervised learning of Reynolds-averaged Navier--Stokes equations. We consider as a test-case the flow past bluff bodies; we train the model using the wake past a cylinder, at different Reynolds numbers 40