Identification of Carbonate Rock Logging Lithology Based on Improved XGBoost Algorithm

The Carboniferous carbonate reservoirs in Kazakhstan’s A Oilfield exhibit significant heterogeneity and a wide variety of rock types. Additionally, the distribution of lithological sample quantities in core data is highly imbalanced, with some lithologies being underrepresented. Traditional lithology identification methods often struggle with such complex reservoirs, leading to challenges such as low accuracy and high labor costs. To address these issues, this study proposes an improved XGBoost algorithm that integrates sample balancing using the Synthetic Minority Oversampling Technique (SMOTE) and parameter optimization through Particle Swarm Optimization (PSO). This method aims to accurately identify six types of limestone in the study area, including sparry oolitic limestone, micritic oolitic limestone, sparry bioclastic limestone, micritic bioclastic limestone, sparry detrital limestone, and micritic detrital limestone. Rock thin sections and their corresponding well-logging data from five wells were used as samples. The SMOTE algorithm was applied to balance the number of samples for each lithology, ensuring equal representation, while the PSO algorithm optimized the model’s parameters to enhance performance. As a result, a carbonate lithology identification model was established. Application results demonstrate that the proposed method outperforms traditional algorithms, such as RF and GBDT, achieving the highest lithology identification accuracy of 92.2%. Furthermore, the improved XGBoost model significantly reduced training time, enhancing overall efficiency. This approach provides a highly accurate and generalizable method for lithology identification in carbonate reservoirs, addressing the challenges posed by complex rock types and imbalanced datasets.