Rock type prediction and 3D modeling of clastic paleokarst fillings in deeply-buried carbonates using the Democratic Neural Networks Association technique

This study outlines a probabilistic model based on artificial neural networks applied to the deeply-buried karsted carbonates of the Ordovician Yingshan Formation, which represent significant oil reservoirs in western China. The complexity of both rock type prediction and 3D facies modeling of paleokarst fillings, which are hosted within the cavities, drives the need to apply innovative techniques for identifying new oil plays. Due to the high heterogeneity of clastic fillings and patchy continuity of the karst patterns, physical evaluation of these reservoirs is extremely complex. We propose the Democratic Neural Networks Association (DNNA) as the probabilistic technique to solve these challenges. This technique simultaneously runs several artificial neural networks in parallel and combines seismic data and well logs. The resulting probable facies volume is expected to provide an appropriate distribution and delineation of clastic fillings (i.e., conglomerates, fine-grained sandstones, siltstones, mudstone, dolomite fragments, and sparry calcarenite) and unfilled or empty spaces. This calculated volume is then used as a reliable input data to condition trend analysis on a very fine geological grid, in order to model the complex patterns in question. The static model obtained shows that, the probabilistic distribution of each filling has the same orientation as karst system. Likewise, spatial dimensions similar to the proposed analogue model of these patterns (vertical and horizontal scales) are delineated. Finally, we validated prediction results by comparing them with the interpreted karst facies of a well not initially considered in the 3D model. The results indicating that the DNNA technique proves to be a useful innovative tool for generating realistic depictions of fillings deposited within deeply-buried paleokarst.