Development of a Machine Learning-Based Conversion Model for Uniaxial Compressive Strength Prediction Between Different Specimen Geometries

Abstract

Uniaxial compressive strength (UCS) is one of the most critical design parameters in rock engineering applications, including the design and construction of engineering structures, underground excavations, and slope stability. Specimens required for UCS testing must be prepared in accordance with various national and international standards. However, when the rock structure is weak or brittle, obtaining the required number and size of specimens may not be feasible. In such cases, alternative specimen geometries and sizes recommended by different standards are often adopted. Although the influence of specimen shape and size on UCS has been extensively studied in the literature and is relatively well understood, the relationship between UCS values obtained from different geometries remains largely unexplored. Notably, there is a lack of comparative studies focusing specifically on rock samples. Furthermore, some transformation equations proposed in the literature have proven inadequate in accurately estimating the strength conversion between cylindrical and cubic specimens. In this study, various machine learning (ML)-based regression algorithms were applied to predict UCS values for cylindrical specimens using UCS values obtained from cubic specimens. A comparative evaluation was conducted using linear regression, tree-based models, ensemble learning methods, kernel-based algorithms, and robust regression techniques. Model performances were assessed through 5-fold cross-validation using R², MAE, MAPE, and RMSE as evaluation metrics. The findings reveal that models such as the Huber Regressor and Support Vector Regression (SVR) provided highly accurate predictions, with narrower error margins and stronger generalization capacity compared to classical transformation coefficients. These results suggest that ML-based models offer an effective and reliable approach offering a robust alternative to conventional transformation equations, especially in engineering contexts where direct experimental testing is limited or impractical.