Decoding Diagenesis with Machine Learning: A New Frontier in Reservoir Diagenesis Evolution Analysis

Summary

Diagenetic processes critically influence the heterogeneity and quality of carbonate reservoirs. With this study, we address the limitations of traditional, subjective methods for quantifying diagenesis by integrating machine learning (ML) with petrological and geochemical analyses. Well log data from 19 wells were analyzed across 15 gas fields on the right bank of the Amu Darya Basin, Turkmenistan. Porosity and permeability data were obtained through routine core analysis on 4,088 core samples. In addition, carbon (C)-oxygen (O) isotope and strontium (Sr) isotope analyses were performed on 57 and 39 core samples, respectively. For ML-based textural quantification, 414 core samples were collected from 17 wells in the Kugitang Formation of the same basin, and 2,260 carbonate photomicrographs were prepared for feature extraction using the fine-tuning segment anything model (SAM). These were combined with logging and core porosity/permeability data to construct a two-step predictive model using Shapley additive explanations (SHAP), particle swarm optimization (PSO), and extreme gradient boosting (XGBoost). The first XGBoost model predicts five carbonate textural components from well log data, while the second XGBoost model integrates these components to predict porosity and permeability, enabling a comprehensive evaluation of diagenetic evolution and reservoir quality. Compared with support vector regression (SVR), random forest regression (RFR), and a hybrid convolutional neural network–long short-term memory (CNN-LSTM) model, the XGBoost approach demonstrates superior predictive accuracy and generalization. For carbonate textural components prediction, it achieves a higher test-set coefficient of determination (R2 = 0.72) than the comparative models (R2 = 0.49 to 0.66). In reservoir property prediction, it also achieves the best performance for porosity (R2 = 0.78) and permeability (R2 = 0.71), outperforming the comparative models (R2 = 0.61 to 0.74 for porosity and 0.53 to 0.64 for permeability). The results highlight the significance of primary porosity in forming high-quality hydraulic flow units (HFUs), particularly HFU-4, within the Kugitang Formation. A total of 23 diagenetic events—comprising early marine cementation, meteoric dissolution (validated by terrigenous-sourced high-radiogenic strontium isotope ratios), and thermochemical sulfate reduction (TSR)—play a crucial role in generating reservoir heterogeneity. Indicators of TSR include negative shift (~7‰) in δ13C (from ~4‰ to approximately −3.1‰), fluid inclusion temperatures exceeding 110°C, and the occurrence of hydrogen sulfide (H2S) and diagenetic pyrite. Thus, the evolutionary model of carbonate reservoirs in the Kugitang Formation is divided into four distinct stages, established by the integration of three complementary data sets: (1) petrographic observations, (2) geochemical analyses, and (3) textural components quantified via ML. The scalable and interpretable workflow bridges traditional methodologies with modern data-driven approaches, providing novel insights into diagenetic processes and prediction for high-quality carbonate reservoirs.