Adaptive Knowledge Fusion Ensemble (AKFE) for Flow Zone Indicator Prediction and Hydraulic Flow Unit Classification from Well Logs

Accurate hydraulic flow characterization in uncored reservoir intervals remains challenging because the Flow Zone Indicator (FZI) requires core-derived porosity and permeability, which are available only at discrete depths. This study introduces an Adaptive Knowledge Fusion Ensemble (AKFE) for continuous FZI prediction and subsequent Hydraulic Flow Unit (HFU) classification from conventional well logs. AKFE integrates well-log features, heterogeneous model predictions, model-disagreement information, and ensemble consensus through a leakage-controlled meta-learning framework. Seven base learners are combined using a CatBoost meta-regressor, and the resulting FZI predictions are classified into four interpreter-defined HFUs using a Gaussian Mixture Model (GMM). The framework was developed using 1,072 depth-matched core observations from two cored wells and independently evaluated on three blind wells. AKFE outperformed Simple Averaging, Weighted Averaging, and Conventional Stacking, achieving an R² of 0.991, RMSE of 0.0374, MAE of 0.0291, and MAPE of 6.7936. Compared with Conventional Stacking, AKFE reduced RMSE, MAE, and MAPE by approximately 26.8%, 29.7%, and 50.0%, respectively. Blind-well application further demonstrated that the predicted FZI profiles preserved meaningful variations in hydraulic quality and yielded consistent HFU patterns supported by cumulative-FZI and porosity–permeability analyses. The results demonstrate that integrating geological context with collective model behavior can improve the transferability of machine-learning-based hydraulic characterization to uncored reservoir intervals.