Lithofacies logging identification for strongly heterogeneous deep-buried reservoirs based on improved Bayesian inversion: The Lower Jurassic sandstone, Central Junggar Basin, China

The strong heterogeneity characteristics of deep-buried clastic low-permeability reservoirs may lead to great risks in hydrocarbon exploration and development, which makes the accurate identification of reservoir lithofacies crucial for improving the obtained exploration results. Due to the very limited core data acquired from deep drilling, lithofacies logging identification has become the most important method for comprehensively obtaining the rock information of deep-buried reservoirs and is a fundamental task for carrying out reservoir characterization and geological modeling. In this study, a machine learning method is introduced to lithofacies logging identification, to explore an accurate lithofacies identification method for deep fluvial-delta sandstone reservoirs with frequent lithofacies changes. Here Sangonghe Formation in the Central Junggar Basin of China is taken as an example. The K-means-based synthetic minority oversampling technique (K-means SMOTE) is employed to solve the problem regarding the imbalanced lithofacies data categories used to calibrate logging data, and a probabilistic calibration method is introduced to correct the likelihood function. To address the situation in which traditional machine learning methods ignore the geological deposition process, we introduce a depositional prior for controlling the vertical spreading process based on a Markov chain and propose an improved Bayesian inversion process for training on the log data to identify lithofacies. The results of a series of experiments show that, compared with the traditional machine learning method, the new method improves the recognition accuracy by 20%, and the predicted petrographic vertical distribution results are consistent with geological constraints. In addition, SMOTE and probabilistic calibration can effectively handle data imbalance problems so that different categories can be adequately learned. Also the introduction of geological prior has a positive impact on the overall distribution, which significantly improves the accuracy and recall rate of the method. According to this comprehensive analysis, the proposed method greatly enhanced the identification of the lithofacies distributions in the Sangonghe Formation. Therefore, this method can provide a tool for logging lithofacies interpretation of deep and strongly heterogeneous clastic reservoirs in fluvial-delta and other depositional environments.