Methods for identifying complex lithologies from log data based on machine learning

The sedimentary environment of the marine-continental transitional facies leads to complex reservoir lithologies characterized by rapid changes in lithology, the presence of thin interbeds, and strong heterogeneity in both vertical and horizontal directions. The resultant difficulties with lithology identification pose challenges to the evaluation of reservoir parameters and the prediction of sweet spots. This study investigated the reservoirs of marine-continental transitional facies in the Permian Longtan Formation in area A, southwestern Sichuan. Based on core descriptions, mineral composition analysis, and log responses, this study divided the reservoir lithologies into eight types, i.e., coals, carbonaceous shales, mudstones, argillaceous siltstones, siltstones, fine sandstones, calcareous mudstones, and bauxitic mudstones. Then, this study determined the log response characteristics of different lithologies, established the charts for identifying typical lithologies from log data, and formed a set of methods and processes for identifying complex reservoir lithologies from log data. Furthermore, this study established lithology identification models using methods of multi-resolution graph-based clustering (MRGC), cross plot - decision tree, and random forest. These models had an accuracy rate of 84.3%, 85.6%, and 91%, respectively, indicating high identification precision overall. In addition, this study compared and analyzed the application conditions, advantages, and disadvantages of different lithology identification methods, providing guidance for subsequent identification of complex reservoir lithologies from log data.