Ensemble machine learning approach for prediction of cementation factor (m) using well logs in heterogeneous reservoirs

The cementation factor ( m ) is a critical parameter in Archie’s equation for water saturation estimation; however, its continuous determination is generally restricted to core-derived measurements. This study presents a robust workflow for continuous prediction of m in heterogeneous mixed-lithology (carbonate–sandstone) reservoirs using well logs and machine learning (ML) techniques. Well log data ( RHOB , NPHI , DT , PHIE , LLD , LLS , and GR ) and core measurements ( φ , F , and m ) were available in two wells (A and B), whereas only well logs were available in three nearby wells (C, D, and E). Feature selection using F-regression and mutual-info-regression identified the most informative logs for predicting the formation resistivity factor ( F ) and m . Nine ML algorithms, including DT, ET, MLP, SVR, LINEAR, RIDGE, KNN, GB, and RF, were trained using Well A and validated on Well B as a blind test well. Ensemble models were subsequently developed using the best-performing algorithms: RF, LINEAR, KNN, GB, ET, and SVR for the F-Ensemble log, and RF, KNN, and GB for the m-Ensemble log. The F-Ensemble achieved R 2 values of 0.98 and 0.86, with RMSE values of 2.13 and 20.22 in Wells A and B, respectively. The m-Ensemble yielded R 2 values of 0.92 and 0.69, with RMSE values of 0.09 and 0.23, respectively. Ensemble-based uncertainty was quantified from the central 10th − 90th percentile spread of the selected models; the normalized mean spreads were 6.51% and 7.26% for F , and 2.6% and 3.2% for m , in Wells A and B, respectively. In addition, an m-Archie log was calculated directly from Archie’s equation using PHIE and the F-Ensemble log, producing reliable agreement with core-derived m values. Comparison with four conventional empirical models demonstrated that the proposed ensemble framework provides superior predictive accuracy, smoother log responses, and greater robustness against noise, outliers, and lithological heterogeneity. The developed methodology was successfully extended to three uncored wells, yielding geologically consistent continuous m profiles. Overall, the proposed workflow provides a reliable approach for continuous estimation of m under limited core-data conditions and demonstrates transferability across nearby wells within the studied heterogeneous carbonate–sandstone reservoir which remaining largely independent of lithology type; however, application to other reservoirs or geological basins would require local recalibration and independent validation.