In modern science, the demand for numerical simulations has significantly increased, particularly in fields such as high-resolution, multi-scale plasma physics and astrophysics. Particle-in-cell (PIC) and hydrodynamic codes (FLASH) are commonly employed to model complex physical phenomena. They rely on solving partial differential equations, which can be extremely time-consuming. The substantial computational cost associated with these simulations has driven a growing interest in developing faster, data-driven surrogate models to expedite traditional simulation processes. Here we propose a novel machine learning approach that addresses the complex micro-physics of transport properties in plasma systems. This study uses Graph Neural Network to represent the continuum system in mesh-based simulations and then apply deep learning techniques to obtain a representation of the underlying physical process. A Physics-Informed Neural Network layer then fine-tunes the predictions by enforcing constraints from mass conservation laws described by nonlinear partial differential equations, enhancing accuracy. The surrogate model can then be applied to replicate the results of traditional simulations at a fraction of the computational expense, reducing their development time and cost. Here we show a preliminary implementation of these techniques to predict evolution of 2D hydrodynamic simulation (Sedov-Taylor explosion).