Perpustakaan Digital ITB

Core-based rock typing is limited by the sparse availability of core data, whereas well logs are recorded continuously along the wellbore. This study evaluates whether unsupervised machine learning applied to well logs can reproduce a core-derived petrophysical rock type reference in the 'X' field, Central Sumatra Basin. A reference was built from 85 routine core analysis (RCAL) samples using the RQI–????z–FZI–DRT framework, and the eight original DRT classes were lumped into four hydraulic-flow-unit (HFU)-derived rock types (RT1 – RT4). Four unsupervised methods: K-Means, FCM, GMM, and MRGC, were applied to five standardized log features (GR, DT, SP, RHOB_STD, and a harmonized deep resistivity proxy, RES_DEEP) from three wells, with all core-derived variables excluded to prevent data leakage. The clusters were mapped using the Hungarian algorithm and evaluated per well and for all wells combined. GMM gave the best combined agreement (accuracy 42.35%, ARI 0.0367), but the per well results were weak (near random in the most sampled well (WS-002) and collapsed to a single rock type in WS-004) indicating that the available conventional logs only partially reproduce the HFU-derived rock types.