Real-time lithology identification based on dynamic felling strategy and differential evolutionary random forest algorithm

Abstract

Lithology identification holds a pivotal role in geological exploration and reservoir characterization, as it directly affects the efficient development and accurate localization of oil and gas resources. However, traditional machine learning approaches often face limitations such as insufficient accuracy and weak generalization when dealing with complex geological conditions. To address these challenges, this study proposes an intelligent lithology identification method that integrates a Differential Evolution (DE) algorithm with a Dynamic Purity Pruning strategy, referred to as DRF-DE. Specifically, the DE algorithm is employed to globally optimize the hyperparameter boundaries of the Random Forest model, enhancing its adaptability to complex data distributions. Subsequently, a dynamic purity pruning mechanism is introduced to eliminate redundant classifiers based on variations in node purity during training, thereby refining the model structure and improving both stability and interpretability. Experimental results on a well-logging dataset from the North Sea oilfield demonstrate that the proposed DRF-DE model achieves an overall classification accuracy of 98.1 % on the test set, while the out-of-bag (OOB) evaluation yields an accuracy of 97.9 %. Compared with conventional machine learning methods, the DRF-DE model shows significant improvements in recognition accuracy and model robustness. Furthermore, the model maintains high performance across various complex geological formations, indicating strong generalization capability and practical applicability. This research not only advances the intelligence of lithology identification but also provides a novel approach and technical support for the automated interpretation of geological data and the efficient development of oil and gas resources.