Explainable Physics-Guided Machine Learning for Modeling and Prediction of Facies-Supported Unconfined Compressive Strength: Implication for CO 2 Storage

Abstract The geomechanical stability of CO₂ storage reservoirs, particularly the Unconfined Compressive Strength (UCS) of caprocks and reservoir formations, is critical for ensuring long-term containment security. Traditional UCS estimation methods, laboratory testing and empirical correlations, are often cost-prohibitive, time-intensive, or insufficiently accurate for heterogeneous subsurface systems. This study introduces a novel Explainable Physics-Guided Machine Learning (PGML) framework that, for the first time, combines facies-specific empirical equations with five advanced machine learning algorithms, AdaBoost, Gradient Boosting (GBT), Random Forest (RF), CatBoost, and Radial Basis Function Neural Network (RBFNN), in a discrepancy learning mode. Rather than predicting UCS directly, the models learn the residual between empirical predictions and laboratory-measured values using conventional well logs (porosity, density, gamma ray, resistivity, slowness, and ultrasonic velocity) from 45 core samples across seven wells in the Zubair Oil Field, Iraq. This physics-anchored approach preserves physical interpretability while capturing complex data-driven patterns. Results demonstrate substantial improvements across all algorithms: PGML increased the coefficient of determination (R²) from 0.10–0.84 (conventional ML) to 0.82–0.98, with the AdaBoost-PGML model achieving the highest accuracy (R² = 0.975, RMSE = 5.9 MPa), a 72% reduction in prediction error compared to conventional AdaBoost (R² = 0.68, RMSE = 21 MPa). SHapley Additive exPlanations (SHAP) analysis reveals that depth, porosity, bulk density, and gamma ray are the primary controls on the empirical-model discrepancy, indicating that compaction trends, pore structure, and clay content, absent from simple velocity-based equations, critically influence UCS. Applied to CO₂ storage assessment in the Zubair Field, the framework identifies a thick sandstone reservoir (φ = 16.9%, k = 428 mD) with high-integrity caprock (UCS = 156 MPa) as highly suitable for secure sequestration, while flagging a deeper interval where moderate caprock strength (UCS < 100 MPa) poses leakage risks. This hybrid PGML-XAI approach provides a robust, interpretable, and practically implementable tool for geomechanical characterization, directly supporting risk-informed decision-making in carbon storage projects.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-12
DOI
https://doi.org/10.21203/rs.3.rs-10080825/v1
Primary Topic
CO2 Sequestration and Geologic Interactions
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Explainable Physics-Guided Machine Learning for Modeling and Prediction of Facies-Supported Unconfined Compressive Strength: Implication for CO 2 Storage

Watheq J. Al‐Mudhafar, Raed H. Allawi, Mohammed Hussein Abdulredha, Abdulsayed H. Abdulsayed et al.
Research Square
CO2 Sequestration and Geologic Interactions
preprint

Explainable Physics-Guided Machine Learning for Modeling and Prediction of Facies-Supported Unconfined Compressive Strength: Implication for CO 2 Storage

Watheq J. Al‐Mudhafar, Raed H. Allawi, Mohammed Hussein Abdulredha, Abdulsayed H. Abdulsayed, Mojtaba M. Hasak, Ali A. Abdulazeez
preprint en

Abstract

Abstract The geomechanical stability of CO₂ storage reservoirs, particularly the Unconfined Compressive Strength (UCS) of caprocks and reservoir formations, is critical for ensuring long-term containment security. Traditional UCS estimation methods, laboratory testing and empirical correlations, are often cost-prohibitive, time-intensive, or insufficiently accurate for heterogeneous subsurface systems. This study introduces a novel Explainable Physics-Guided Machine Learning (PGML) framework that, for the first time, combines facies-specific empirical equations with five advanced machine learning algorithms, AdaBoost, Gradient Boosting (GBT), Random Forest (RF), CatBoost, and Radial Basis Function Neural Network (RBFNN), in a discrepancy learning mode. Rather than predicting UCS directly, the models learn the residual between empirical predictions and laboratory-measured values using conventional well logs (porosity, density, gamma ray, resistivity, slowness, and ultrasonic velocity) from 45 core samples across seven wells in the Zubair Oil Field, Iraq. This physics-anchored approach preserves physical interpretability while capturing complex data-driven patterns. Results demonstrate substantial improvements across all algorithms: PGML increased the coefficient of determination (R²) from 0.10–0.84 (conventional ML) to 0.82–0.98, with the AdaBoost-PGML model achieving the highest accuracy (R² = 0.975, RMSE = 5.9 MPa), a 72% reduction in prediction error compared to conventional AdaBoost (R² = 0.68, RMSE = 21 MPa). SHapley Additive exPlanations (SHAP) analysis reveals that depth, porosity, bulk density, and gamma ray are the primary controls on the empirical-model discrepancy, indicating that compaction trends, pore structure, and clay content, absent from simple velocity-based equations, critically influence UCS. Applied to CO₂ storage assessment in the Zubair Field, the framework identifies a thick sandstone reservoir (φ = 16.9%, k = 428 mD) with high-integrity caprock (UCS = 156 MPa) as highly suitable for secure sequestration, while flagging a deeper interval where moderate caprock strength (UCS < 100 MPa) poses leakage risks. This hybrid PGML-XAI approach provides a robust, interpretable, and practically implementable tool for geomechanical characterization, directly supporting risk-informed decision-making in carbon storage projects.

Research Square
University of Basrah (IQ)
Climate action
CO2 Sequestration and Geologic Interactions
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon