<![CDATA[Electrofacies clustering is fundamental to reservoir characterization but is often hindered by the subjectivity and inefficiency of conventional manual interpretation, particularly in heterogeneous fields. This study presents a robust, data-driven workflow for automating electrofacies identification using unsupervised machine learning, applied to Gamma Ray (GR) logs from 66 wells across 16 reservoir intervals in the NVS Field, Central Sumatra Basin. The methodology systematically evaluates the impact of feature representation by comparing Statistical, Convolutional Neural Network (CNN), and Long Short-Term Memory (LSTM) extraction techniques coupled with K-Means, BIRCH, and Gaussian Mixture Model (GMM) clustering algorithms. Performance assessment using Silhouette scores and the Davies-Bouldin Index demonstrates that LSTM-based features consistently yield superior clustering results by capturing critical sequential log-shape dependencies essential for resolving vertical heterogeneity. While algorithmic efficacy was found to be context-dependent—with GMM favoring transitional facies and K-Means excelling in high-contrast zones—the integration of the optimal models successfully reconstructed geological patterns without prior labeling. External validation against reference facies maps confirmed that the unsupervised clusters exhibit strong spatial coherence, accurately delineating the Northwest-Southeast (NW-SE) depositional trend of Tidal Bar Axis and Margin zones. Furthermore, the model demonstrated high geological sensitivity by successfully identifying localized features such as Isolated Sand Bars. These findings verify the geological plausibility of the proposed workflow and underscore the necessity of sequence-aware feature extraction, offering a reproducible and objective framework for reservoir modeling in data-limited environments.]]>
