Physics-informed neural network (PINN) is an alternative technique to solve partial differential equations (PDEs). The PINN approach can solve any dimension of ordinary and/or partial differential equations through automatic differentiation incorporating initial and boundary conditions. PINN loosely satisfies the local mass balance through a loss function; however, there is no comprehensive study on the deviation of balance laws. In this study, we developed a PINN framework to solve PDEs for steady-state groundwater flow in both homogeneous and heterogeneous porous media. Next, we investigated the local and global mass balances of the PINN solution. Finally, we discussed the local and global mass balances of PINN through numerical h-convergence, iteration, and collocation points.