The construction of subsurface reservoir models is typically aided by the use of outcrops and modern analogue systems. We show how process-based models of depositional systems help to develop and substantiate reservoir architectural concepts. Process-based models can simulate assumptions relating to the physical processes influencing sedimentary deposition, accumulation and erosion on the resultant 3D sediment distribution. In this manner, a complete suite of analogue geometries can be produced by implementing different sets of boundary conditions based on hypotheses of depositional controls. Simulations are therefore not driven by a desired/defined outcome in the depositional patterns, but their application to date in reservoir modelling workflows has been limited because they cannot be conditioned to data such as well logs or seismic information.
In this study a reservoir modelling methodology is presented that addresses this problem using a two-step approach: process-based models producing 3D sediment distributions that are subsequently used to generate training images for multi-point geostatistics.
The approach has been tested on a dataset derived from a well-exposed outcrop from central Utah. The Ferron Sandstone Member includes a shallow-marine deltaic interval that has been digitally mapped using a high-resolution unmanned aerial vehicle (UAV) survey in 3D to produce a virtual outcrop (VO). The VO was used as the basis to build a semi-deterministic outcrop reference model (ORM) against which to compare the results of the combined process/multiple-point statistics (MPS) geostatistical realizations. Models were compared statically and dynamically through flow simulation.
When used with a dense well dataset, the MPS realizations struggle to account for the high levels of non-stationarity inherent in the depositional system that are captured in the process-based training image. When trends are extracted from the outcrop analogue and used to condition the simulation, the geologically realistic geometries and spatial relationships from the process-based models are directly imparted onto the modelling domain, whilst simultaneously allowing the facies models to be conditioned to subsurface data.
When sense-checked against preserved analogues, this approach reproduces more realistic architectures than traditional, more stochastic techniques.