Implications of rootless geothermal models: Missing processes, parameter compensation, and imposter convection

•Models of heat and mass in geothermal reservoirs commonly represent the upper part a deep convecting system and simplify the deep region to a boundary.•All choices of boundary condition can misrepresent aspects of the overall system, including essential processes like convection.•Parameters can take unphysical values to compensate for incorrect physics, affecting model calibration and biasing forecasts.•Forecast models must account for these simplifications. Numerical models of geothermal systems commonly capture only the top of a reservoir. Deeper areas of the reservoir are simplified to a boundary at the base of the model domain. Commonly, the basal boundary is given either a heat source or a source of mass and enthalpy. We developed and present simple numerical experiments which demonstrate that these approaches do not produce the correct model behavior in comparison to a model that captures the entire convecting domain with a heat flux only. We describe a variety of incorrect types of model behavior that arise directly from the choice of boundary condition, independent of the specific parameterization of the model. Heat sources are sensitive to the thickness of the domain and parameters take unphysical values to compensate for the reduced height. The combined mass/heat boundary can produce temperatures that show similarities to circulating geothermal systems, but with incorrect fluid flow and a strong boundary layer focused at the base of the simulated clay cap. These errors likely cause parameters to adjust their values to compensate for the incorrect physics. We highlight these issues and show an example from a developed reservoir model. Initial calibrations to natural state temperature were unsuitable for history matching. The 3D model required parameter adjustments to achieve a more realistic production model. Appropriate mitigation measures should be considered to reduce parameter compensation and improve decision-support models.