Explainable Machine Learning for Tight Sandstone Reservoirs: A Zonal Shap and Domain-Guided Eda Framework for Productivity Classification

Gharavi, Amir, O'Sullivan, Aidan, Haddad, Malik, Hassan-Sayed, Mohamed, Yosefi, Paria and Al-Saegh, Salam (2026) Explainable Machine Learning for Tight Sandstone Reservoirs: A Zonal Shap and Domain-Guided Eda Framework for Productivity Classification. Geoenergy Science and Engineering, 262 (214436).

Abstract

Traditional physics-based models often struggle to characterize tight sandstone formations due to complex pore geometries and limited fluid connectivity. To address these limitations, this study introduces an explainable machine learning (ML) framework tailored for predicting and interpreting reservoir productivity in highly heterogeneous systems. The framework integrates domain-informed exploratory data analysis (EDA), principal component analysis (PCA), and SHapley Additive exPlanations (SHAP), combined with a Random Forest (RF) classifier, to ensure both predictive accuracy and geological interpretability. The target variable—SuperRT, a binary indicator of reservoir productivity—was defined using integrated petrophysical thresholds and engineering judgement. To capture vertical heterogeneity, the dataset was restructured into three GeoZones (A, B, and C) based on facies-specific reservoir quality. Key features used for modeling include effective porosity (PHIE), oil saturation (SO), permeability (Perm), shale volume (VSH), and total organic carbon (TOC). Zone-specific SHAP analysis revealed distinct productivity drivers: Zones A and C were dominated by fluid-driven properties (SO, Perm), while Zone B was controlled by storage and source characteristics (PHIE, TOC), highlighting differing stimulation requirements. The RF model achieved strong predictive performance with an F1-score of 0.77, recall of 0.83 for productive zones, and cross-validation accuracy of 91.0% ± 2.3%. By embedding explainability at both global and zone levels, this study bridges the gap between data science and subsurface domain knowledge. The proposed workflow enhances transparency, supports operational decision-making, and offers a field-ready, interpretable solution for development planning in unconventional reservoirs.

Actions (login required)

Edit Item Edit Item
Altmetric