Multidimensional Adaptive Deep Clustering for Intelligent Diagenetic Facies Logging Recognition

Summary Diagenetic facies are indicators that characterize the nature, type, and quality of reservoirs. Identifying diagenetic facies is critical for efficient exploration and development of tight sandstone gas reservoirs. Currently, research on diagenetic facies recognition focuses on the use of supervised learning methods based on labeled data sets. However, the high cost of generating diagenetic facies labels, due to the use of sophisticated instrumentation and expert annotation, limits the generalization of supervised learning methods. Therefore, there is an urgent need for an efficient and adaptive unsupervised learning method in the field of diagenetic facies recognition. In unsupervised learning, diagenetic phase identification faces the problems of complex data distribution and optimizing the number of clusters. To address these issues, we propose a multidimensional adaptive deep clustering model (MADELINE). The advantages of this model are as follows: (1) To address the problem of low-quality feature extraction caused by differences in data distribution, we propose an asymmetric autoencoder module (ASAE) that uses internal features of log data to extract representative multidimensional features of attributes. (2) Design the adaptive cluster number learning module (ACME) to realize adaptive learning of clustering parameters and solve the problem of intelligent optimization of clustering parameters. Log data from six wells (AC, CNL, DEN, GR, SP, RT) in the Ordos Basin and six representative models were selected for comparative experiments. The results show that MADELINE achieves optimal performance in both the Silhouette Coefficient (SC) and the Calinski-Harabasz Index (CHI). The proposed method improves the SC metric by 8.94% compared to classic clustering and deep clustering algorithms. To further verify the effectiveness of MADELINE, ablation experiments, hyperparameter analysis, and visualization analysis of the experimental results were performed. The results of the ablation experiments show that each module of MADELINE plays a positive role, and the asymmetric model structure is more beneficial for feature learning. The MADELINE model proposed in this study effectively solves the problems of data distribution differences and adaptive optimization of cluster numbers in diagenetic facies recognition and provides a low-cost and rapid diagenetic facies recognition method for oil exploration and development. This research has practical application value for reservoir quality evaluation, oil-bearing prediction, and exploration and development decision-making.