Enhancing Lithology Prediction through Preprocessing Techniques and Machine Learning Models

Abstract

Lithology classification through well log interpretation is a fundamental task in reservoir characterization, enabling accurate delineation of subsurface formations and assessment of hydrocarbon potential. However, measurements are rarely full, and missing data intervals are prevalent due to operational difficulties or logging device failure. Thus, imputation of missing data from down-hole well logs is a prevalent issue in subsurface processes. Our work a strong emphasis on the preprocessing phase and data imputation, acknowledging that missing data in well logging is a prevalent problem that can have a major impact on classification results. Our work is part of the FORCE2020 Lithology Classification Competition. Our method underlines how important extensive data preprocessing is for improving model performance, including regression-based imputation, normalization, and class balancing by SMOTE. Traditional models like Random Forest and XGBoost were able to produce reliable results in the challenging FORCE2020 Lithology Classification. By leveraging multiple models, we aim to enhance the accuracy and robustness of our predictions, addressing the challenges posed by missing data and ensuring a more reliable classification process. We show that the Random Forest model obtains the greatest accuracy of 95% using the FORCE 2020 dataset from 118 wells in the Norwegian Sea. This study emphasizes how crucial thorough data imputation and preprocessing techniques are to raising the precision and dependability of lithology classification.