Explainable machine learning and deep learning for productive zone identification in tight sandstone reservoirs: integrating PROMETHEE-II and class imbalance handling

Abstract

Identifying productive zones in tight sandstone reservoirs is hampered by geological heterogeneity, severe class imbalance, and the need for transparent model interpretation. We present an integrated, explainable machine learning and deep learning workflow that combines zonal analysis and multi-criteria model selection via the Preference Ranking Organization METHod for Enrichment Evaluations (PROMETHEE-II). Applied to a West African tight sandstone dataset, the proposed approach achieved outstanding predictive accuracy (F1-score = 0.95, recall = 1.00, ROC-AUC = 1.00). While these values indicate excellent discrimination capability for the present dataset, they may partly reflect dataset characteristics, stratified cross-validation structure, and effective class-imbalance correction. Therefore, model uncertainty and the potential for performance variability across other reservoirs should be considered when generalising these results.
The workflow also provided robust, geologically validated insights using zonal SHapley Additive exPlanations (SHAP). Core-derived flow unit and network connectivity analyses further confirmed the reliability of the predicted productive intervals. By delivering a transparent, “glass-box” solution, this framework improves both reservoir productivity classification and operational decision-making for targeted hydraulic stimulation in unconventional resources.

More information

Identifiers

ORCID for Mohamed G. Hassan: orcid.org/0000-0003-3729-4543

Catalogue record

Export record

Altmetrics