Prediction of Uniaxial Compressive Strength Using Fully Bayesian Gaussian Process Regression (fB-GPR) with Model Class Selection.

In rock, mining, and/or tunneling engineering, determination of uniaxial compressive strength (UCS) of rocks is an important and crucial task, which is often estimated from readily available index properties of rocks in practice, such as porosity (n), Schmidt hammer rebound number (Rn), P-wave velocity (Vp), point load index (Is(50)). This is especially true for projects with medium- or small-size, as well as for rocks with a high degree of fragility and porosity. While numerous methods have been proposed for predicting UCS indirectly, linear assumptions are frequently made during model development, despite the possibility of nonlinear relationships between UCS and the abovementioned indices. Furthermore, many established methods often struggle to strike a balance between model complexity and performance, resulting in models that are either over- or under-fitted. As a result, constructing an optimal UCS model with minimal variables while maintaining a high level of performance remains a great challenge in rock engineering practice. This paper proposes a fully Bayesian Gaussian process regression (fB-GPR) approach to develop an optimal model for UCS prediction which strikes a balance between prediction accuracy and model complexity. Both real-world and numerical examples are used to illustrate the proposed method. Results show that the optimal model of predicting UCS for the database from Malaysia is constructed by only n and Vp, with the same coefficient of determination of around 0.9 as the more complex model involving n, Rn, Vp and Is(50). A sensitivity study is also performed to systematically examine its robustness and accuracy of the proposed method in developing optimal model for UCS prediction. Highlights: A fully Bayesian Gaussian process regression (fB-GPR) method is proposed for constructing an optimal model for UCS prediction from rock indices. The optimal model for UCS prediction has minimal variables but maintains a high level of performance. The fB-GPR approach is data-driven and non-parametric, and can automatically strike a balance between model complexity and performance. [ABSTRACT FROM AUTHOR]