Enhancing Hoek–Brown constant estimation accuracy based on rigidity index using deep neural network
摘要
Estimation of the Hoek–Brown failure criterion constant (mi) is crucial for geotechnical design but remains challenging due to inherent nonlinearity in rock mass behavior. Existing empirical equations for mi prediction exhibit limitations, particularly at lower threshold values of the rigidity index. This study aims to develop an optimized deep learning approach to minimize mi estimation errors. To develop a model for predicting mi, a training dataset comprising literature-sourced data on various intact rock types was compiled. Statistical analyses characterized the parameter distributions and relationships. Eight empirical formulations were also implemented to estimate mi values. Subsequently, four predictive models of varying complexity were designed and trained utilizing machine learning and deep neural network techniques. The models took various combinations of uniaxial compressive and indirect tensile strengths as inputs. Rigorous performance evaluation demonstrated deep learning formulations attained perfect fits and consistently outperformed empirical relationships. Notably, incorporating rigidity index resolved issue in characterizing mi below threshold values. This study, therefore, validates artificial intelligence’s ability to systematically enhance rock failure criterion characterization by resolving known challenges.