This study outlines a probabilistic model based on artificial neural networks applied to the very deep karsted carbonates of the Ordovician Yingshan Formation, which represent significanct reservoirs within a region of the Tahe oilfield, Tarim Basin, China. The complexity of rock type prediction and distribution of paleokarst fillings hosted in cavities, drives the need to apply new techniques for identifying more plays. This investigation focuses on a karsted interval located between the reflections of unconformities T74 and 76. The analysis was conducted using acoustic impedance (P-wave) and amplitude seismic attributes, processed from a 32-bit seismic dataset (Poststack). The methodology also includes conventional wireline logs from 28 wells adjusted to lithological descriptions of cores. Democratic Neural Networks Association (DNNA) is the proposed method for rock type prediction in karsted carbonates that simultaneity utilizes 3-D seismic data and well data. Based on sedimentological descriptions, the karst facies are classified in six types of lithofacies: mixed siliciclastic and carbonate (e.g., calcarenite and conglomerate), limestone, very fine-grained sandstone, mudstone, breccia, and unfilled. According to identified lithofacies, a clustering analysis is performed using the followings logs: Gamma Ray (GR), Deep Resistivity (RD), Neutron (CNL), Density (DEN), Sonic (AC), Potassium (K) and Thorium (TH) from Spectral Gamma Ray. Subsequently, the outputs are simplified for selecting the model most representative of lithofacies. Once adjusted data in commercial package, a training set and stabilization geometry involving seismic attributes constrained by interpreted seismic horizons are processed. The extracted seismic traces along borehole trajectory after processing demonstrate a good match with analyzed data, where the predicted maximum probability and class probability tracks vary with respect to lithofacies. Making time slices on the computed volume are observed to estimate effectively the probability of karst facies away from the wells. The outcome of this workflow is a probabilistic facies volume that provides appropiate description of clastic rocks that cover paleokarst fillings essentially in the run-off subzone. The model indicates that mudstone facies are the most prevailing and better proportions of siltstones or sandstones facies are distributed to southeast of area. This study concludes that in determining clastic lithofacies distribution, employing several neural networks running in parallel that simultaneously learn from the same dataset through different strategies is an effective tool. To date, there has not been any study on rock type prediction using DNNA in karsted carbonates. The results represent a significant contribution to the collection of geosciences on characterizing this type of reservoir.