Research and Application of Rock Typing Using Deep Learning in Prediction of Carbonate Reservoirs of H Oilfield, Iraq

The loss of some important logging data especially lithology data in old oilfields and huge cost in manpower and material on drilling and coring have brought great difficulties to the development of oilfields. A new method using deep learning for rock typing is achieved to classify and count the limited logging data, establish appropriate pore-permeability (PP) relationship and reduce the risk of reservoir prediction, which provides a concise and effective way for carbonate rock prediction. In allusion to the existed problems, the paper collects, ranks the correlation between rock types and conventional logging data, which establishes a neural network model based on deep learning, divides the carbonate reservoirs into 4 types, and estimates the pore-permeability relationship for each type. Finally, a pore-permeability cloud simulation was performed based on the geo-statistical inversion to set up a high-precision reservoir static model with perfect well-seismic tie. The reliable permeability property can be obtained which helps to accurately depict the spatial distribution of the reservoirs. The carbonate reservoir of M formation for H oilfield in the Middle East is of complex pore structure with strong heterogeneity and poor relationship of the pore-permeability (PP). The logs DT, GR, Density, Porosity as the input features of deep learning is optimized to train neural network models, which are applied to the sample for testing and verification. The sample tests from the optimized neural network model are as accurate as 86.8%. The results show that rock typing using deep learning and well logs found a non-linear mapping relationship which effectively and reasonably divided the carbonate reservoirs into 4 types. The geological statistics and stochastic simulation in geo-statistical inversion organically combine seismic information with rich reservoir parameters such as porosity, rock typing, permeability and so forth. The permeability inversion result is highly consistent with the drilling data, which means the reservoir distribution patterns and regularity have been greatly improved, the local characteristics of the reservoir have been described more detailed and accurate. The paper establishes and optimizes a neural network model based on deep learning which extends the divided rock typing to the whole oilfield. The estimated permeability as new pore-permeability relationship was applied to the geo-statistical inversion, which achieved the high-resolution spatial prediction of reservoir parameters and satisfied the fine reservoir characterization. It reduces the huge cost on drilling and coring, also provides a concise and effective approach to improve the reservoir estimation and production efficiency.