Abstract Rock strength parameters, including uniaxial compressive strength (UCS), tensile strength (TS), cohesion (c), and angle of internal friction (φ), are essential to evaluate the stability conditions of the ground in civil, mining, petroleum, and geological engineering. Determination of these parameters is time-consuming, expensive, destructive, lab-oriented, and sample-intensive. With the advancement of machine learning technology, numerous prediction models were proposed by previous researchers, though the role of lithological control remained relatively unexplored. To explore the lithological control and the performances of different machine learning models such as support vector regression (SVR), K-nearest neighbor regression (KNN), random forest regression (RFR), XGBoost regression (XGBR), CatBoost regression (CBR), AdaBoost regression (ABR), gradient boosting regression (GBR), and light gradient boosting regression (LGBR) were employed to predict the UCS, TS, c, and φ using the P-wave velocity (VP) and lithology information, which are easy, cheap, non-destructive, and can be obtained in the field. The performance of the models was evaluated and compared for lithology (L) and without-lithology (WL) information using root mean squared error (RMSE), coefficient of determination (R2), and mean absolute percentage error (MAPE). Overall, the KNN algorithm performed best for the developed lithology-based model with R2 values of 0.9976, 0.9973, 0.9957, and 0.9838 for UCS, TS, c, and φ, respectively. Finally, Shapley Additive Explanations (SHAP) and feature importance analyses were performed using the XGBR model for individual prediction of VP, UCS, TS, c, and φ with different combinations of input parameters. Research highlights <UnorderedList Mark="Bullet"> <ItemContent> Eight machine learning models are used to predict the geomechanical properties (UCS, TS, c, and φ of rock materials using a multi-output regression method. </ItemContent> <ItemContent> Considered rock type as an input parameter to establish lithological control by comparing both lithological and non-lithological prediction models. </ItemContent> <ItemContent> The prediction performance of five boosting machine learning models is compared for geomechanical property prediction which is not reported before to our best knowledge. </ItemContent> <ItemContent> Prediction mechanism and influence of rock type on each parameter (UCS, TS, c, and φ) is established using SHAP analysis. </ItemContent> </UnorderedList>

Lithological control on machine learning-based prediction of rock strength parameters using P-wave velocity

期刊论文
发表日期：

2026年6月26日

查看全文

Kripamoy Sarkar,
Tabish Rahman,
Shyam Raj B,
Trilok Nath Singh

摘要

Abstract

Rock strength parameters, including uniaxial compressive strength (UCS), tensile strength (TS), cohesion (c), and angle of internal friction (φ), are essential to evaluate the stability conditions of the ground in civil, mining, petroleum, and geological engineering. Determination of these parameters is time-consuming, expensive, destructive, lab-oriented, and sample-intensive. With the advancement of machine learning technology, numerous prediction models were proposed by previous researchers, though the role of lithological control remained relatively unexplored. To explore the lithological control and the performances of different machine learning models such as support vector regression (SVR), K-nearest neighbor regression (KNN), random forest regression (RFR), XGBoost regression (XGBR), CatBoost regression (CBR), AdaBoost regression (ABR), gradient boosting regression (GBR), and light gradient boosting regression (LGBR) were employed to predict the UCS, TS, c, and φ using the P-wave velocity (V_P) and lithology information, which are easy, cheap, non-destructive, and can be obtained in the field. The performance of the models was evaluated and compared for lithology (L) and without-lithology (WL) information using root mean squared error (RMSE), coefficient of determination (R²), and mean absolute percentage error (MAPE). Overall, the KNN algorithm performed best for the developed lithology-based model with R² values of 0.9976, 0.9973, 0.9957, and 0.9838 for UCS, TS, c, and φ, respectively. Finally, Shapley Additive Explanations (SHAP) and feature importance analyses were performed using the XGBR model for individual prediction of V_P, UCS, TS, c, and φ with different combinations of input parameters.

Research highlights

Eight machine learning models are used to predict the geomechanical properties (UCS, TS, c, and φ of rock materials using a multi-output regression method.

Considered rock type as an input parameter to establish lithological control by comparing both lithological and non-lithological prediction models.

The prediction performance of five boosting machine learning models is compared for geomechanical property prediction which is not reported before to our best knowledge.

Prediction mechanism and influence of rock type on each parameter (UCS, TS, c, and φ) is established using SHAP analysis.