Research and application of logging lithology identification for igneous reservoirs based on deep learning

Abstract

Igneous reservoirs are characterized by heterogeneity and anisotropy, which makes logging interpretation difficult. In order to identify their lithology, deep learning to establish the deep belief network (DBN) by logging data is proposed in this study. By the least square method, the mean square error function is used to measure network performance, and network parameters such as the number of RBMs, the number of neurons in the hidden layer of each RBM, and the classification boundary. Then, the logging data that require interpretation are processed by DBN that is trained. The results are divided into four cases and are analyzed and discussed further. First, the lithology classification results are continuous, stable and the formations are thick. At this point, there is no need to correct the results. Second, there are several lithological discontinuities in the thick layer. In this case, if the thickness of the discontinuous formations is >0.5 m, the corresponding formations can be divided according to the identification results; if the thickness of the discontinuous formations are less than or equal to 0.5 m, the discontinuous formations are merged into adjacent thick formations. Third, the lithology of formations cannot be determined by the identification result. At this time, it is generally considered that the lithology of the formations do not appear in the training samples. Fourth, there are a few identification results for one formation. At this point, a cross plot is adopted to correct these results. An accuracy of 94.8% is achieved for lithology identification by the deep belief network with lithology correction.