Meta ensemble learning-based triaxial rock dynamic strength model

Triaxial dynamic strength is of interest to various fields of engineering and science. The determination of rock strength is a fundamental element of any design and analysis in geomechanics and geoengineering. Data-oriented machine learning (ML) algorithms have been gaining more traction in this field due to their high performance and flexibility. However, an understanding of the capabilities of these paradigms to provide fast, cheap, and accurate predictions of triaxial rock dynamic strength is yet lacking. This study aims to contribute to the field of rock dynamics by employing two stacking and voting ensemble methods and four ML algorithms, namely Gaussian process (GP), random forest (RF), decision table (DT), and K-nearest neighbor (KNN) for modeling the dynamic triaxial strength of rock material. A database of 267 experiments compiled from available published laboratory triaxial tests on seven rock materials was used for the development of the ensemble models. The triaxial tests were carried out under different confining pressures and strain rates. Therefore, the input variables in these models are rock type, confining pressure (up to 450 MPa), and strain rate (ranging from \(10^{ - 8}\) ${10}^{-8}$ to 600 \({\text{s}}^{ - 1}\) ${\text{s}}^{-1}$ ), with the output being the major principal stress. Based on the results, RF, KNN, voting, and stacking models performed better than GP-RBF, GP-PUK, and DT models in terms of accuracy and error metrics in the training and testing datasets. This indicates that the approaches used are capable of capturing the dynamic triaxial strength of rock material. A parametric study using the cosine amplitude method indicates that confining pressure, rock type, and strain rate are the most to least effective variables on the responses of tests in all evolved surrogate data-driven models. This study also aims to address the gap in the literature concerning prediction data-driven surrogate models in triaxial rock dynamic strength criteria and related subfields.