Filling-well: An effective technique to handle incomplete well-log data for lithology classification using machine...

Area of Science:

• Geoscience and Petroleum Engineering
• Machine Learning Applications in Earth Sciences
• Data Science for Resource Exploration

Background:

• Lithology classification is vital for sustainable oil and gas exploration.
• Missing values in well-log data (e.g., GR, NPHI, RHOB, RS, DTCO, DTSM, RD) significantly reduce machine learning classification accuracy.
• Addressing missing data is critical for reliable subsurface characterization.

Purpose of the Study:

• To evaluate machine learning algorithms for handling missing values in well-log datasets.
• To improve lithology classification accuracy in the presence of significant data gaps (up to 30%).
• To compare the performance of XGBoost, KNN, and ANN in imputing missing well-log data.

Main Methods:

• Application of Extreme Gradient Boosting (XGBoost) algorithm.
• Implementation of K-nearest neighbours (KNN) algorithm.
• Utilization of Artificial Neural Network (ANN) algorithm, including preprocessing with isolation forest and bias correction.

Main Results:

• XGBoost demonstrated the highest efficiency and accuracy, achieving the lowest Mean Absolute Percentage Error (MAPE) and Root Mean Square Error (RMSE) for RHOB, NPHI, DTCO, and DTSM.
• ANN performed well on GR, RS, and RD features post-preprocessing but showed potential for overfitting with smaller datasets.
• KNN exhibited limitations with missing-not-at-random (MNAR) data due to its reliance on distance metrics and the k parameter.

Conclusions:

• XGBoost is the most effective algorithm for handling extreme missing values in well-log data for lithology classification.
• ANN offers a viable alternative, particularly for specific log types, but requires careful tuning and sufficient data.
• KNN is less suitable for MNAR well-log data imputation compared to XGBoost and ANN.