Evaluation of machine learning methods for lithology classification of sandstone-type uranium deposit based on well logging data in the Songliao Basin, Northeast China

Lithology identification plays a crucial role in uranium exploration, providing important information about geological conditions, deposit characteristics, and mineralization environments, which is significant for the exploration, development, and management of uranium resources. Traditional lithology identification methods have issues such as low precision, slow recognition speed, and poor generalization performance. Artificial intelligence, with the advancement and development of science and technology, has provided a new research direction for uranium ore bed lithology identification. By combining the multivariate parameter well logging data of sandstone-type uranium mines in the Songliao Basin with machine learning algorithms, models such as Random Forest (RF), Support Vector Machine (SVM), Multi-Layer Perceptron (MLP), and LightGBM (Light Gradient Boosting Machine) have been established to conduct comprehensive lithology identification from well logging data to achieve the purpose of tapping the potential of well logging data. The research shows that after training on sample wells in the study area, all four models can accurately predict the lithology of more than 89% of the samples. This shows that machine learning methods are useful for figuring out the lithology of uranium ore beds. Among the four test wells, the LightGBM model achieved the highest lithology identification accuracy of up to 98.8%, followed by the RF and MLP models with accuracies of 97.9% and 96.0%, respectively, while the SVM model performed relatively worse, with the highest accuracy of mere 90.1%. Moreover, the LightGBM model has to some extent overcome the issue of an imbalanced class distribution of well logging data samples within the study area, providing a reference for the complex lithology identification of sandstone-type uranium strata.