Current reservoir simulators solve Advection-Diffusion-Reaction equations on unstructured grids to a high degree of accuracy but are computationally expensive to use in history matching, optimization, and uncertainty quantification workflows. We propose a novel machine learning architecture to auto-regressively predict state variables and well outputs of a reservoir for an ensemble of cases with varying well configurations. This new approach constructs a Graph Neural Network (GNN) model to predict hyperbolic variables (i.e., saturations and compositions), a Fourier Neural Operator (FNO) model to predict elliptic variables (i.e., pressure), and a feed-forward layer for well outputs. Predictions from the FNO model are fed into the GNN and are coupled in an auto-regressive manner to predict an arbitrary future state of the system given just the initial state of the reservoir. The outputs of the model at each timestep are then fed into the well output layer to predict outputs for each injector and producer well in the reservoir. Results are compared with solutions from a high-fidelity reservoir simulator applied to an ensemble workflow.