Lithofacies division and intelligent identification of the lacustrine mixed rocks in the Upper Xiaganchaigou Formation in Yingxi area of the Qaidam Basin, northwestern China

The Lower Ganchaigou Formation in the Yingxi area of the Qaidam Basin is a typical lacustrine mixed rock reservoir in western China. It is characterized by strong interlayer heterogeneity, development of diverse lithofacies types, and complex response features in logging curves. These complexities make lithofacies identification of the Ganchaigou Formation particularly challenging for non-coring wells, demanding a more efficient and accurate approach. Based on lithology and structural patterns, a lithofacies classification scheme was established. Three intelligent logging identification methods based on improved long short-term memory (LSTM) networks were constructed for lithofacies identification. The accuracy of these methods was evaluated, and the most suitable intelligent logging identification method for the reservoir lithofacies in the Yingxi area was selected. In the Upper Xiaganchaigou Formation (E 3 2 section) of the Yingxi area, a total of eight lithofacies types were identified: laminated lime-dolostone, stratified lime-dolostone, laminated dolostone-lime, stratified dolostone-lime, laminated lime-dolomitic shale, massive mudstone, sandstone, and gypsum. The overall recognition accuracies of the LSTM, Bi-LSTM, and Attention-based Bi-LSTM intelligent identification models are 81%, 85%, and 87%, respectively. The overall recognition accuracies of the three intelligent algorithms are relatively high, with the Attention-based Bi-LSTM model achieving the highest accuracy. This model demonstrates superior applicability for intelligent lithofacies identification in lacustrine mixed rock reservoirs, particularly those dominated by carbonates in the Yingxi area. It effectively interprets the lithofacies types of non-coring wells in the study area and provides a valuable reference for interpreting lithofacies logs in similar depositional environments. • Based on lithology and structural patterns, eight lithofacies types were identified in the study area. The depositional characteristics and logging responses of these lithofacies are significantly different, making the classification scheme practical. • An Attention-based Bi-LSTM model was developed by incorporating the attention mechanism, serving as the intelligent lithofacies identification algorithm. This architecture enables the model to effectively process well logging sequence data, targeting the strong heterogeneity between lacustrine mixed sedimentary rock layers. It can identify the logging response of complex lithological changes, highlight key features, and better achieve intelligent prediction. • The Attention-based Bi-LSTM model has higher recognition accuracy and is more suitable for intelligent lithofacies identification in the carbonate-dominated lacustrine mixed rock reservoirs of the Yingxi area.