Effective Machine-Learning Models for Rock Mass Deformation Modulus Estimation Based on Rock Mass Classification Systems

Mohammad Khajehzadeh1,2

Suraparb Keawsawasvong1,Email

Mohammad Reza Motahari3

Pitthaya Jamsawang4
 

1Research Unit in Sciences and Innovative Technologies for Civil Engineering Infrastructures, Department of Civil Engineering, Thammasat School of Engineering, Thammasat University, Pathumthani, 12120, Thailand.
2Department of Civil Engineering, Anar Branch, Islamic Azad University, Anar 1477893855, Iran.
3Department of Civil Engineering, Faculty of Engineering, Arak University, Arak PO Box 379, Iran.
4Soil Engineering Research Center, Department of Civil Engineering, King Mongkut’s University of Technology North Bangkok, Bangkok 10800, Thailand.

Abstract

The rock mass deformation modulus (RMDM) plays a crucial role in dam and tunnel design. This study introduces advanced machine-learning (ML) models to predict RMDM using rock mass rating (RMR) and the Q-system at the Khersan-2 dam site in southwestern Iran. Through the analysis of exploratory boreholes, the engineering geological properties of the samples, Q, RMR, RMDM, geological strength index (GSI), Hoek-Brown, and shear strength constants of the rock mass were determined. Subsequently, seven effective ML models, namely random forest, multilayer perceptron backpropagation artificial neural network, Gaussian process regression, K-nearest neighbor, simple regression, and multiple linear and non-linear regression approaches, were utilized to estimate RMDM. Based on classification systems, the site was rated as having good RMR and Q categories. A new empirical relationship with high accuracy was established between Q and RMR89. Furthermore, RMDM demonstrated a strong correlation with Q and RMR, as supported by statistical analysis. The results showed the relative superiority of non-linear regression models compared to linear ones. The employed ML techniques displayed remarkable accuracy in estimating RMDM, achieving a coefficient of determination (R2) greater than 97%. Notably, Gaussian process regression with a squared exponential kernel function stood out as the most effective approach, yielding outstanding performance in predicting RMDM with an impressive R2=0.99 and RMSE=0.01 compared to all other investigated methods.

Effective Machine-Learning Models for Rock Mass Deformation Modulus Estimation Based on Rock Mass Classification Systems