3D training image development and conditioning strategies for multiple-point statistical simulations

In Multiple-Point Statistical (MPS) approaches, the training image (TI) and the conditioning data play a crucial role (Mariethoz and Caers, 2014). In fact, MPS combines the ability to condition the realizations to hard and soft data with the capability to reproduce geological features characterized by statistical properties formalized via the TI. In the present research, we compare different strategies for SNESIM simulations on a large 3D model volume. The simulation domain (∼45·106voxels) corresponds to the Miocene formation in the south of Denmark (Figures 1a-b). The Miocene sediments can be roughly subdivided into two lithologies: sand and clay. An already existing geological model (the Tønder model, Figure 1b) is present in part of the study area. A portion of the Tønder model is utilized as starting point for the iterative development of the 3D TI. Subsequently, the TI is adjusted based on the a-posteriori analysis of the associated unconstrained realizations (Figure 2). In addition, the seismic lines in the area are used as hard conditioning information as their reliability is higher than the other available data (Figure 1c). Hard conditioning is also used to ensure a perfect correspondence between the simulation results and the pre-existing Tønder model (Figure 1c). On the contrary, because of their general lower quality, and different discretization (1m) compared to the size of the realization grid (5m), the borehole data are included as soft probability (Figure 3b). Actually, in order to avoid the limitations of SNESIM, and successfully reproduce the spatial trend in the clay/sand ratio across the investigated domain, we find it effective to interpolate the sand probability derived from the boreholes into a 3D voxel model and use it as soft conditioning (Figures 3c-d). In summary, this research shows a possible workflow to properly build TIs and effectively handle input information to be successfully used for large-scale geostatistical 3D modelling (Figures 4d-f).