Perpustakaan Digital ITB

Reservoir characterization is a fundamental step in resource development and production. This step involves identifying rock types to infer key reservoir parameters such as permeability. Traditional methods rely heavily on core data, which is often limited and expensive. To address this limitation, this study enhances the abundant log data with machine learning techniques to predict porosity, permeability, and rock type across the entire well depth. The methodology involves comprehensive data preprocessing, including integrating, normalizing, cleaning Log ASCII Standard (LAS) and RCAL data and feature extraction. Rock typing was performed using RQI/FZI methods, and the data was split into training and testing sets. Specifically, data from six wells with core data were used. One well was designated as the testing set, and the remaining wells were used for training. This process was repeated for each well in a loop, ensuring that all wells were evaluated individually. Three main types of machine learning models were assessed: regression, classification, and clustering algorithms. Sequential forward selection (SFS) was employed to identify the optimal features for the models, and cross-validation was used to ensure their robustness. The models provided good predictions despite the limited data (103 data points). Results indicated that regression models for porosity prediction achieved an overall RMSE of 3.66%, demonstrating reasonable accuracy. For rock-type prediction, the clustering model achieved an average accuracy of 46.6% across all wells, outperforming both regression and classification models. Permeability predictions further validated the clustering model's superiority, with the lowest RMSE values averaging 42.572 md across most wells. Compared to the manual rock-typing average RMSE of 44.83 md, machine learning proved its reliability in generalizing to new data. This study underscores the potential of machine learning to integrate abundant log data with core data, enhancing subsurface property predictions. The clustering model proved the most effective, demonstrating robust performance and reliable predictions across different datasets. These findings highlight the viability of machine learning as a powerful tool in petrophysics and reservoir engineering, offering significant advantages over traditional core data analysis methods.