Exploration of lithology identification technology based on generative adversarial networks

The lithology serves as the fundamental component for reservoir assessment and plays a pivotal role in the calculation of reservoir parameters, evaluation, and development. Well logging data contains crucial formation information, which forms the basis for lithology identification. However, due to various influencing factors, traditional methods often result in multiple solutions in lithology interpretation, leading to reduced accuracy. Leveraging the advantages of machine learning in data analysis and modeling, this study employs a stepwise lithology identification strategy and generative adversarial neural network to identify logging lithology with unbalanced categories. Based on well logging data from Permian and Triassic sandstone reservoirs in Fukang Depression, this paper analyzes the correlation between well logs and petrophysical analysis. Seven logging curves including acoustic time difference (AC), Caliper (CALI), neutron porosity (CNL), density (DEN), natural gamma ray (GR), formation resistivity (RT) and SP were selected as characteristic inputs for identifying six types of lithologies with good recognition effect: mudstone, sandy mudstone, fine sandstone, medium sandstone, sand conglomerate and conglomerate. Comparative tests show that multi-step lithology identification has an accuracy 4.21% higher than single-step identification. Furthermore, the two-step adophytic network model demonstrates an accuracy 4.72%-7.19% higher than RF, SVM, XGBoost, and LSTM models. The overall identification accuracy reaches 83.44%, indicating promising prospects for its application in the field of lithology identification.