Ensemble modeling is important to account for the uncertainty in the underlying physical processes. Although exhaustive, numerical simulation can be computationally expensive when running multiple realizations. Representing the reservoir models using simple graphs that connect well completions is an emerging technique to decrease such computational costs. In this work, we present a geologically informed graph that is history-matched within the simulator framework and results in accurate and fast surrogate models. Methods presented in the literature create graphs by connecting wellbores using straight lines and then converting the resulting topology to a numerical simulation grid. These generally work well for cases with no prior reservoir model but naturally ignore stratigraphic information. In this work, we present an embedded method within the numerical simulator. Once the wellbore data is loaded, a graph-based grid is created using structural and geological information, if available. Once the connections are identified, fastest path methods such as Dijkstra's algorithm are used to build the computational graph. Resulting models represent a reduced order grid in terms of the numerical control volume, reducing both CPU run-time and memory consumption for reservoir simulations, allowing the usage of massive parallelization in history matching workflows.