Fast multiple-point simulation using a data-driven path and an efficient gradient-based search

Multiple-point geostatistics has recently attracted significant attention for modeling different environmental variables. These methods employ the patterns of a training image (TI) to complete a simulation grid (SG), resulting in realizations with good spatial continuity and structural properties. Most existing multiple-point statistics (MPS) methods scan the SG in a random or raster order. In this paper, a new method is presented with a data-driven scanning path giving high priority to pixels with high gradient magnitude. As a result, the image edges are synthesized first, resulting in better connectivity preservation. Although MPS methods usually produce promising results compared to traditional variogram-based modeling, their further development is somehow limited by their excessive computational burden. An efficient search space reduction method, consistent with the proposed ordering scheme, is also presented in this paper. Experiments on different geological fields show results comparable to the state-of-the-art with a significant improvement in CPU time. Graphical abstractDisplay Omitted HighlightsAttentively selecting the scanning order to improve continuity of patterns.Expanding the attentive path selection idea to 3D grids.Giving more priority to pixels with more hard neighbors to improve conditioning.Reducing the search space using gradient information to speed up the simulation.