Evaluation of rock mechanical properties is crucial in various engineering applications, including oil and gas development, underground gas storage, and CO₂ geological storage. This study utilized nanoindentation, scanning electron microscopy, and energy-dispersive spectroscopy (SEM–EDS) analyses to build a comprehensive dataset encompassing shale material type, content, distribution characteristics, and micromechanical properties. The dataset was then used to develop XGBoost models that capture the coupling relationships between microstructures and mechanical properties, enabling accurate estimation of hardness (H) and reduced modulus (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">${E}_{r}$</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mi>E</mi> <mi>r</mi> </msub> </math></EquationSource> </InlineEquation>) at microscopic scale. The results indicate that the predicted H and <InlineEquation ID="IEq2"> <EquationSource Format="TEX">${E}_{r}$</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mi>E</mi> <mi>r</mi> </msub> </math></EquationSource> </InlineEquation> values closely match the measured values, with absolute errors typically below 1 GPa and 10 GPa, and relative errors generally less than 20%. On test dataset, the coefficient of determination (<InlineEquation ID="IEq3"> <EquationSource Format="TEX">${R}^{2*}$</EquationSource> <EquationSource Format="MATHML"><math> <mmultiscripts> <mrow> <mi>R</mi> </mrow> <mrow /> <mrow> <mn>2</mn> <mrow /> <mo>∗</mo> </mrow> </mmultiscripts> </math></EquationSource> </InlineEquation>) of predicted and measured H, <InlineEquation ID="IEq4"> <EquationSource Format="TEX">${E}_{r}$</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mi>E</mi> <mi>r</mi> </msub> </math></EquationSource> </InlineEquation> are higher than 0.80. The model demonstrates strong generalization ability and performs well in practical applications. It enables the evaluation of shale micromechanical properties from discrete points to continuous surfaces, with predictions aligning closely with actual measurements. Combined with an upscaling method (Mori–Tanaka method), macroscopic mechanical parameters—Young’s modulus and Poisson's ratio—can be approximated based on the predicted micromechanical properties. In summary, the dataset and XGBoost model developed in this study enable accurate estimation of shale mechanical properties, providing critical mechanical data for energy-related engineering and offering a methodological reference for other materials.Highlights<UnorderedList Mark="Bullet"> <ItemContent> Developing a novel big dataset of microscopic rock mechanical parameters. </ItemContent> <ItemContent> Constructing machine learning models for predicting microscopic rock mechanical properties. </ItemContent> <ItemContent> Estimation of shale micro and macro mechanical parameters from high–resolution SEM–EDS images. </ItemContent> </UnorderedList>

A Machine Learning Approach for Estimating Shale Micro-Mechanical Properties from High Resolution SEM–EDS Images

期刊论文
发表日期：

2025年8月14日

查看全文

Yang Chen,
Shuheng Tang,
Zhaodong Xi,
Shasha Sun,
Pengfei Jiao,
Xiaofan Mei,
Tianqi Zhou,
Jian Sun,
Fanhui Zeng

摘要

Evaluation of rock mechanical properties is crucial in various engineering applications, including oil and gas development, underground gas storage, and CO₂ geological storage. This study utilized nanoindentation, scanning electron microscopy, and energy-dispersive spectroscopy (SEM–EDS) analyses to build a comprehensive dataset encompassing shale material type, content, distribution characteristics, and micromechanical properties. The dataset was then used to develop XGBoost models that capture the coupling relationships between microstructures and mechanical properties, enabling accurate estimation of hardness (H) and reduced modulus ( ${E}_{r}$ $E_{r}$ ) at microscopic scale. The results indicate that the predicted H and ${E}_{r}$ $E_{r}$ values closely match the measured values, with absolute errors typically below 1 GPa and 10 GPa, and relative errors generally less than 20%. On test dataset, the coefficient of determination ( ${R}^{2*}$ $R^{2 *}$ ) of predicted and measured H, ${E}_{r}$ $E_{r}$ are higher than 0.80. The model demonstrates strong generalization ability and performs well in practical applications. It enables the evaluation of shale micromechanical properties from discrete points to continuous surfaces, with predictions aligning closely with actual measurements. Combined with an upscaling method (Mori–Tanaka method), macroscopic mechanical parameters—Young’s modulus and Poisson's ratio—can be approximated based on the predicted micromechanical properties. In summary, the dataset and XGBoost model developed in this study enable accurate estimation of shale mechanical properties, providing critical mechanical data for energy-related engineering and offering a methodological reference for other materials.

Highlights

Developing a novel big dataset of microscopic rock mechanical parameters.

Constructing machine learning models for predicting microscopic rock mechanical properties.

Estimation of shale micro and macro mechanical parameters from high–resolution SEM–EDS images.