This thesis investigates the use of Graph Neural Networks (GNNs) as surrogate models for computational fluid dynamics problems characterized by fundamentally different physical and mathematical properties. The objective is to evaluate whether the same message-passing architecture, trained separately on different datasets, can effectively model fluid flows governed by distinct physical assumptions and mathematical properties. In the first case a convergent–divergent nozzle governed by the 2D compressible Euler equations is considered. A parametric dataset is generated by varying the chamber-to-exit pressure ratio, leading to different flow regimes ranging from fully subsonic to mixed elliptic–hyperbolic behavior with shock formation. The GNN is trained to perform node-wise regression of Mach number and static pressure fields.