Perpustakaan Digital ITB

Formation prediction is important for understanding subsurface structures and making better reservoir models. Traditionally, geologists define formation boundaries using core samples and cutting descriptions. However, this method can be subjective and may lead to inconsistent results between wells. To solve this, machine learning is used in this study to classify formations based on well log data. Two types of machine learning methods are applied: supervised and unsupervised. Supervised learning uses labeled data to train a model, while unsupervised learning looks for patterns in data without labels. Five well log features (GR, NPHI, RHOB, PEF, and DT) are used in supervised models, and three features (GR, NPHI, RHOB) are used in unsupervised models. The Random Forest model achieved the highest F1-score (0.9090), followed by XGBoost and Decision Tree. Agglomerative Clustering gave the best result among unsupervised methods with a silhouette score of 0.5243. These results show that supervised learning is better for formation prediction, while unsupervised learning helps to understand hidden patterns in the data.