<![CDATA[This study applies a machine learning approach to classify lithology using well log data from 14 wells in Ford County, Kansas, United States, to address the limitations of conventional interpretation, which is time-consuming and subjective due to overlapping log responses. Reference lithology labels were generated using predefined well-log interpretation criteria and grouped into four classes: sandstone, limestone, shale/clay, and coal. Two supervised learning algorithms, K-Nearest Neighbors (KNN) and Artificial Neural Networks (ANN), were evaluated and compared. The preprocessing stages included data cleaning by removing null values and inconsistencies, Z-score normalization, class balancing using SMOTE on the training data to prevent data leakage, and feature selection based on Pearson correlation. Model performance was evaluated using Classification Accuracy (CA), Area Under the Curve (AUC), Logarithmic Loss (Log Loss), and 5-fold cross-validation. The results indicate that ANN consistently outperformed KNN in lithology classification. ANN achieved classification accuracies above 95%, AUC values approaching 1.00, and low Log Loss, whereas KNN achieved testing accuracies of approximately 75-80% but exhibited lower cross-validation performance, indicating reduced robustness in intervals characterized by overlapping lithological responses. The optimal ANN architecture consisted of three hidden layers with 100-100-100 neurons and 100 training iterations. Visual evaluation of four test wells showed good agreement between the predicted and reference lithology distributions. These findings suggest that machine learning, combined with appropriate preprocessing techniques, can support lithology classification from well log data. Among the evaluated models, ANN demonstrated superior capability in capturing nonlinear relationships between well log responses and lithological variations within the study area.]]>
