Diagenetic Facies Identification in the Santos Basin Pre‐Salt Carbonates Using Geologically Guided Feature Engineering and Ensemble Machine Learning

ABSTRACT

Silicification and dissolution in the Santos Basin pre‐salt lacustrine carbonates produce overlapping wireline‐log responses that complicate diagenetic‐facies identification. The problem is amplified by strong class imbalance and pseudo‐replication arising from the 0.00254 m export interval. Fixed interpretation templates can, therefore, be unreliable, particularly for the rare composite silicification–dissolution facies. A geology‐guided sample‐refinement stage combined a weighted coefficient of variation with class‐specific feature‐space distance thresholds. It reduced an intermediate training table from 1.22 million to 90,061 records (92.6%) by preferentially retaining heterogeneous intervals, thin dissolution beds, and most silicification and composite records. Twenty‐four physics‐driven attributes were then combined with nine retained original logs, and a four‐stage selection procedure reduced 33 candidate variables to 11 predictors. Random forest (RF), LightGBM, and CatBoost were compared using fivefold well‐grouped cross‐validation and two independent blind wells, so samples from one well did not enter both training and validation folds. RF performed best under the present validation design, with blind‐well macro‐ F 1 scores of 0.78 and 0.74, dissolution recall of 78.6% and 93.4%, and composite recall of 99.9% and 100.0%. Composite precision was more moderate (38.6% and 52.1%), so these predictions are best interpreted as high‐sensitivity screening rather than definitive facies boundaries. The geology‐guided workflow outperformed both raw‐data RF and optimized SMOTE baselines. SHAP attribution patterns were consistent with the expected mineralogical and pore‐system responses, although they do not independently establish causality. The principal contribution is a transferable workflow rather than a new classifier: Geological knowledge is incorporated into label construction, sample refinement, feature engineering, model interpretation, and well‐level validation. The resulting probability and facies logs can support reservoir screening, interwell correlation, and uncertainty‐aware upscaling, subject to local geological calibration.