This research aims to significantly advance the understanding of crack stress thresholds
in rocks by utilizing intelligent and statistical methodologies applied to a comprehensive
database of granitic rocks. The database includes essential mechanical parameters,
such as crack initiation stress and crack damage stress, as outputs, and unconfined
compressive stress, Young's modulus, and Poisson's ratio as inputs. The study employs
multivariate regression analysis (MRA), Decision tree (DT), and artificial neural
network (ANN) techniques to analyze the data. To assess and compare the performance
of these models, various metrics are used, including root-mean-square error (RMSE),
mean absolute error (MAE), coefficient of determination (R²), mean absolute percent
error (MAPE) and the a20-index. The findings reveal that the ANN method surpasses
MRA-based and decision tree regression-based approaches in accurately predicting the
output parameters. The ANN model demonstrates outstanding predictive performance for
σci, achieving an R²=0.934, RMSE=5.965, MAE=4.79, MAPE=5.3% and VAF=93.34%. Similarly,
the model excels in predicting σcd with an R²=0.985, RMSE=4.676, MAE=3.785, MAPE=2.3%,
VAF=98.4%, Additionally, sensitivity analysis identifies unconfined Compressive stress
(UCS) as the most critical factor in predicting crack initiation stress (σci) and
crack damage stress (σcd). The findings suggest that the ANN model is more effective
in predicting σci and σcd of rocks compared to other methods.
