Journal Article

Identification of diagenetic facies based on data difference enhancement (DDE) machine learning

Ye, Tairan, Xu, Qin, Bi, Youyi, Xu, Caihua, Zhu, Peng, Ma, Tong, Feng, Ying

2025

Overview

Abstract Accurate identification of diagenetic facies is crucial for reservoir characterization, as it directly determines the evaluation of petrophysical properties, pore structure types, and reservoir quality, thereby playing a pivotal role in predicting high-quality hydrocarbon-bearing zones. However, conventional identification approaches are often limited by subjective interpretation, lack of standardized criteria, and low operational efficiency. Meanwhile, existing intelligent classification techniques frequently fail to adequately discriminate subtle but critical data variations, leading to suboptimal classification accuracy. To address these challenges, this paper develops a novel method for diagenetic facies identification based on data difference enhancement (DDE). The proposed methodology consists of three key steps: first, high-dimensional well-logging data are projected into a lower-dimensional feature space using the t-SNE algorithm to improve computational efficiency while preserving nonlinear relationships. Subsequently, the k-means clustering algorithm partitions the processed dataset into distinct groups, thereby amplifying intra-cluster data homogeneity and inter-cluster separability. Next, an ensemble learning architecture is constructed using the stacking algorithm, where cluster-specific meta-classifiers are individually optimized to enhance model robustness. During application, unclassified samples are assigned to their nearest cluster based on Euclidean distance metrics, followed by targeted prediction using the corresponding meta-classifier. The data from the second member of the Upper Triassic Xujiahe Formation are employed for model evaluation, and the results show that the DDE machine learning method significantly outperforms conventional machine learning algorithms, including k-nearest neighbours, support vector machines, and random forests, achieving an accuracy rate of 86.4%. This workflow enables efficient and reliable diagenetic facies classification using standard well-logging curves, offering both theoretical insights and practical tools for reservoir quality prediction in hydrocarbon exploration.

• Publisher: Oxford University Press
• Subject: Algorithms / Classification / DDE / Hydrocarbons / Machine learning / Nitrous oxide / Reservoirs / Triassic