Abstract: The progression into a phase of rock tunnel construction is marked by significant dimensions,notable length,substantial depth,and pronounced complexity.Uncertain geological conditions within surrounding rock formations,combined with limited expert resources,give rise to myriad challenges during the excavation process of the New Austrian Tunnelling Method(NATM) tunnel construction,primarily concerning rock mass quality assessment and excavation safety evaluation.This results in a sequence of scientific inquiries,encompassing aspects related to rock mass structural attribute characterization,hierarchical modeling,and methodologies for safety assessment during excavation.The present study revolves around these scientific inquiries,employing a diverse range of methodologies including on-site measurements,statistical data analysis,intelligent algorithms,and numerical simulations.An algorithm has been devised for quantitatively extracting characteristics of the rock mass face,concomitantly establishing a refined hierarchical model for the classification of rock masses by integrating a variety of heterogeneous data sources.Building upon this foundational work,an investigation into the safety evaluation of tunnel excavation within complex geological environments has been executed.The principal achievements of this study are outlined as follows:Addressing the challenges posed by soft interlayers,joint fissures,and subterranean aquifers has resulted in the development of databases for image semantic segmentation and apparent structure image classification.The application of deep learning algorithms and methodologies such as hyperparameter optimization has facilitated accurate classification and nuanced representation of feature information.Consequently,a comprehensive 13-dimensional heterogeneous database has been established,encapsulating geometric,environmental,and physico-mechanical parameters pertinent to the rock mass.Through the construction of a hybrid TPE-GBRT prediction model,the optimal parameter combination for predictive optimization within the hybrid machine learning model has been identified,facilitating precise prediction of rock mass classification as indicated by the RMR index.Lastly,the geological environment of the rock mass has been modeled on the basis of a discrete fracture network(DFN),leading to the formulation of a three-dimensional tunnel model within 3DEC,grounded in excavation surface data.The continuous tunnel excavation process has been simulated using strength reduction methods,enabling the assessment of stability characteristics,including stress-strain responses,shear displacement,and safety conditions.