Klasifikasi Litologi Berbasis Machine Learning Pada Data Well Log

Garini, Sherly Ardhya (2025) Klasifikasi Litologi Berbasis Machine Learning Pada Data Well Log. Doctoral thesis, Institut Teknologi Sepuluh Nopember.

Text 7025222003-Doctoral.pdf - Accepted Version Restricted to Repository staff only Download (17MB) | Request a copy

Abstract

Litologi merupakan jenis batuan penyusun reservoir migas yang memiliki pori dan permeabilitas untuk mengalirkan fluida. Klasifikasi litologi penting dilakukan baik pada reservoir lama maupun baru untuk mendukung eksplorasi energi dan kemandirian energi nasional. Metodologi klasifikasi litologi berbasis machine learning telah banyak dikembangkan, namun kualitas data well log yang tidak lengkap menjadi tantangan utama. Penyebab umum hilangnya data log adalah kerusakan alat, kondisi lubang (washout), dan kesalahan penyimpanan. Tujuan penelitian ini adalah untuk mengevaluasi kinerja dan ketahanan metodologi klasifikasi litologi berbasis machine learning terhadap kehilangan data pada well log, serta menilai performa berbagai metode penanganan missing values. Berdasarkan hasil penelitian yang mengacu pada tiga tujuan utama, dapat disimpulkan bahwa Algoritma Light Gradient Boosting Machine (LGBM) menunjukkan kinerja terbaik baik pada kondisi data lengkap (baseline) maupun saat data mengalami kehilangan nilai. Pada eksperimen baseline menggunakan data aktual DTCO, LGBM mencapai akurasi 98,1% dan F1-Score 98,2%, menegaskan kemampuannya dalam menangkap hubungan non-linear antar parameter log. Evaluasi terhadap berbagai metode imputasi K-Nearest Neighbors (KNN), Extreme Gradient Boosting (XGBoost), LGBM, Artificial Neural Network (ANN), dan Bidirectional Recurrent Neural Network (BiRNN) menunjukkan bahwa LGBM memberikan peningkatan akurasi dan penurunan error paling signifikan, dengan kinerja MAPE 6,53 kali lebih baik dan RMSE 55,17 kali lebih rendah dibandingkan XGBoost, sehingga direkomendasikan sebagai metodologi utama untuk post-imputation pada data well log tidak lengkap. Selanjutnya, kombinasi LGBM Imputation dan LGBM Classifier terbukti paling stabil dalam menangani Missing Not At Random (MNAR) pada berbagai tingkat masking (5-50%), mempertahankan performa tinggi (penurunan hanya ±1,5% pada F1-Score). Sebaliknya, metode imputasi sederhana seperti mean imputation dan listwise deletion mengalami penurunan signifikan (F1-Score hingga 0,85 dan 0,70) akibat hilangnya variasi geologi alami. Secara geofisika kesalahan klasifikasi yang muncul mencerminkan zona transisi dan kemiripan komposisi mineral antar litologi, sehingga menunjukkan bahwa batas antar batuan bersifat gradual. Dengan demikian, penelitian ini memberikan kontribusi signifikan terhadap pengembangan metode cerdas untuk rekonstruksi dan klasifikasi litologi berbasis data well log.
=================================================================================================================================
Lithology refers to the types of rocks that compose hydrocarbon reservoirs and possess porosity and permeability that enable fluid flow. Lithology classification is essential for both mature and newly developed reservoirs to support energy exploration and national energy independence. Although machine-learning-based lithology classification methodologies have been widely developed, the quality of incomplete well-log data remains a major challenge. Common causes of missing log data include tool failure, borehole instability (washout), and data storage errors. The objective of this research is to evaluate the performance and robustness of machine-learning-based lithology classification methodologies under missing well-log conditions, and to assess the effectiveness of various missing-value handling techniques. Based on findings aligned with the three primary research objectives, the Light Gradient Boosting Machine (LGBM) algorithm demonstrated the best performance both on complete baseline data and under missing-value conditions. In the baseline experiment using actual DTCO data, LGBM achieved an accuracy of 98.1% and an F1-Score of 98.2%, confirming its ability to capture non-linear relationships among log parameters. Evaluation of several imputation methods K-Nearest Neighbors (KNN), Extreme Gradient Boosting (XGBoost), LGBM, Artificial Neural Network (ANN), and Bidirectional Recurrent Neural Network (BiRNN), showed that LGBM produced the greatest improvement in accuracy and the largest reduction in error, achieving MAPE performance 6.53 times better and RMSE 55.17 times lower than XGBoost. Therefore, LGBM is recommended as the primary methodology for post-imputation of incomplete well-log data. Furthermore, the combination of LGBM Imputation and LGBM Classifier proved the most stable in handling Missing Not At Random (MNAR) conditions across masking levels of 5–50%, maintaining high performance with only ±1.5% reduction in F1-Score. In contrast, simple imputation methods such as mean imputation and listwise deletion experienced significant degradation (F1-Score dropping to 0.85 and 0.70) due to the loss of natural geological variability. From a geophysical perspective, misclassification patterns reflect transitional zones and similarities in mineral composition across lithologies, indicating that rock boundaries are gradual rather than discrete. Thus, this study provides a significant contribution to the development of intelligent methods for well-log data reconstruction and lithology classification.

Item Type:	Thesis (Doctoral)
Uncontrolled Keywords:	Klasifikasi, litologi, machine learning, missing values, well log Classification, lithology, machine learning, missing values, well log.
Subjects:	Q Science Q Science > Q Science (General) Q Science > Q Science (General) > Q325.5 Machine learning. Support vector machines.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55001-(S3) PhD Thesis (Comp Science)
Depositing User:	Sherly Ardhya Garini
Date Deposited:	22 Jan 2026 02:47
Last Modified:	22 Jan 2026 02:47
URI:	http://repository.its.ac.id/id/eprint/130035

Actions (login required)

View Item