Perpustakaan Digital ITB

Shallow, thin-bedded gas reservoirs may contain productive intervals that are not identified by conventional deterministic petrophysical cutoffs. This study developed a retrospective machine learning workflow to screen for possible missed pay zones in the Q-Sandstones of the A12-FA Field in the Dutch North Sea. The dataset comprised nine production wells and six raw LWD curves: GR, RHOB, TNPH, DRHO, P16H, and P40H. The curves were processed on their native 0.1524 m grid without interpolation or imputation. Conventional pay was defined in Interactive Petrophysics using PHIE ? 0.3, VSH ? 0.245, and Sw ? 0.623, and thirty-four predictor-safe features were derived from the raw curves. Each well was evaluated through well-held-out cross-fitting using models developed exclusively from the remaining eight wells. Track A assessed the ability of Logistic Regression, Random Forest, and XGBoost to reproduce the conventional PayFlag. XGBoost achieved the highest performance, with macro-well precision of 0.9667, recall of 0.9569, F1-score of 0.9603, and MCC of 0.6646. Track B ranked non-pay samples according to their similarity to Grade-A flow-supported intervals and generated the candidate intervals. The primary 2.0 m criterion identified four intervals in A03, comprising 131 samples and 19.9644 m. The 0.5 m thin-bed sensitivity, implemented as four native samples or 0.6096 m, identified seven intervals, comprising 155 samples and 23.6220 m. All seven intervals occur within the gross Q4 flowing interval. However, the available reports do not allocate flow to individual beds, and the intervals therefore remain unresolved rather than selectively confirmed missed pay. Increasing the Sw cutoff to 0.69 recovered six intervals, corresponding to 88 samples and 13.4112 m, but also classified 255 additional samples outside the candidates as pay. The workflow therefore provides a structured basis for prioritising intervals for petrophysical review. Any cutoff revision or operational decision still requires depth-resolved evidence, such as production logging or selective testing.