What does this research mean for the field?

A feedforward deep neural network trained on high-resolution, laboratory-calibrated petrophysical data can accurately predict the static shear modulus of evaporitic caprocks, outperforming other machine learning models such as CNN, XGBoost, and LightGBM. Novelty: ClaimNovelty.METHODOLOGICAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

What question did this study set out to answer?

The study aims to develop a framework integrating geomechanical data and machine learning to predict the shear modulus of caprock, a critical factor for CO2 and H2 storage resilience.

June 9, 2026Open Access

Application of Petrophysics-based Machine Learning for Caprock Resilience in CO2 and H2 Storage

Key Points

The study aims to develop a framework integrating geomechanical data and machine learning to predict the shear modulus of caprock, a critical factor for CO2 and H2 storage resilience.
Employs a calibration workflow using dynamic elastic moduli from dipole sonic logs to derive static shear modulus labels.
Conducts laboratory tests (triaxial and uniaxial compression) on core samples to calibrate models at depths of 1200 to 1900 m.
Utilizes multiple machine learning models (DNN, CNN, XGBoost, LightGBM) and performs sensitivity analysis for feature contribution evaluation.
Feedforward deep neural network (DNN) achieves R2 of 0.9774, MAE of 0.0352, RMSE of 0.9734 on the test set.
DNN outperforms CNN (R2 0.9745), XGBoost (R2 0.9621), and LightGBM (R2 0.9485).
Sensitivity analysis shows key features like Vs, Vp, and RHOB significantly contribute to predictive performance.

Abstract

A framework is presented that integrates laboratory calibrated geomechanical data with machine learning (ML) to predict the shear modulus (Gs) of caprock as a resilience metric for subsurface CO 2 and H 2 storage. Data driven approaches have been applied to related geomechanical properties in storage contexts; however, a specific gap is addressed through the use of high resolution 10 cm static Gs labels derived from dipole sonic log calibration in evaporitic caprocks, combined with systematic benchmarking and interpretability analysis across multiple models. A transformation and calibration workflow is employed, in which high resolution dynamic elastic moduli obtained from dipole sonic logs, including compressional wave velocity (Vp) and shear wave velocity (Vs), along with bulk density (RHOB), are converted into laboratory calibrated static moduli. Calibration is performed using triaxial and uniaxial compression tests on core samples, following international society for rock mechanics standards, from an anhydrite rich caprock interval at depths of about 1200 to 1900 m under representative confining pressures. This procedure generates 10 cm resolution ground truth static Gs labels for ML training and validation. A comprehensive benchmarking and interpretability analysis is then conducted across multiple models, including feedforward deep neural networks (DNN), one dimensional convolutional neural networks (CNN), Extreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LightGBM). A sensitivity analysis is applied to evaluate the contribution of input features, including Vs, Vp, RHOB, gamma ray (GR), and neutron porosity (NPHI), to model performance and predictive robustness, which improves interpretability and confidence in the developed models. Among the evaluated models, the feedforward DNN achieves the highest predictive accuracy on the test set with R 2 of 0.9774, MAE of 0.0352, and RMSE of 0.9734, and it outperforms CNN with R 2 of 0.9745, XGBoost with R 2 of 0.9621, and LightGBM with R 2 of 0.9485. Complementary metrics, including MAE, RMSE, and MARE, confirm the superior balance between accuracy and generalization achieved by the DNN.

Read Full Paperexternally

AI에게 질문

Bookmark

View Full Paper