Conventional tunnel disaster prevention systems are constrained by several critical limitations, such as inade-quate accuracy in geological detection, oversimplified modeling, delayed simulation responses, subjective risk assessment, and coarse decision-making. These constraints hinder comprehensive situational awareness and sci-entific disaster control in complex tunnel environments. Artificial intelligence (AI) offers essential support for the digital-intelligent upgrading of tunnel disaster prevention by enabling automated interpretation of predic-tive data, multi-source information modeling, rapid disaster scenario simulation, and scientifically-grounded risk evaluation. This study systematically investigates digital-intelligent methods for tunnel disaster prevention and safety protection following the technical chain of “geological prediction-modeling-simulation-assessment and decision-making.” For geological information prediction, research focuses on improving the accuracy and effi-ciency of data analysis through algorithmic optimization to achieve rapid and reliable perception under complex conditions. In modeling, machine learning-based intelligent modeling approaches and digital twin (DT)-driven cyber-physical integration are discussed. Regarding disaster simulation, deep-learning-based surrogate models are summarized for their applications in disaster forecasting, effectively overcoming the high computational cost and response delay inherent in conventional simulation methods. In risk assessment and decision-making, the applications of machine learning (ML) and emerging large language models (LLMs) are examined, highlight-ing advances in risk identification, prediction, and reasoning. Despite these advancements, challenges persist, including strong reliance on limited data sources, lack of physical constraints and interpretability, insufficient generalization across scenarios, and restricted real-time capability. Future developments are expected to pursue deeper integration of data-driven and physics-informed approaches, improve the utilization of multi-source het-erogeneous data and scenario generalization, enhance model robustness and interpretability, and incorporate lightweight architectures with dynamic updating mechanisms, thereby enabling real-time perception, full-chain simulation, and intelligent risk management throughout the entire tunnel engineering processes.

The rapid economic growth and urbanization, in China have led to a substantial expansion of urban and under-ground confined spaces, consequently heightening fire risks. Efficient and environmentally sustainable fire and smoke control in these environments presents significant challenges to public safety and the stable, sustainable extraction of mineral resources. This paper provides a systematic review and summary of three key aspects: the behavioral characteristics of fire smoke in urban and underground confined spaces, technologies for fire detec-tion and early warning, and smoke control techniques. It aims to offer a rational reference for advancing fire emergency prevention and control strategies in China’s urban and underground confined spaces. By addressing the complex interactions between internal and external factors that influence fire smoke behavior in these envi-ronments, this study highlights persistent issues such as monitoring blind spots and false alarms caused by sensor drift. To tackle these challenges, the following key directions are proposed: (1) Developing multi-scale coupled models to deepen research on smoke behavior characteristics in complex environments;(2) Establishing an un-manned, precision early warning and inspection platform based on intelligent sensing and rapid response systems; (3) Creating an intelligent evacuation system utilizing virtual reality (VR), Internet of Things (IoT), and smart sensing technologies; (4) Advancing integrated fire and smoke synchronization control technology to enhance fire prevention and control capabilities in urban and underground confined spaces.

Accurate determination of initial in-situ stress is essential for tunnel support design and safety evaluation. How-ever, direct measurement methods are often restricted by geological conditions and economic factors. To ad-dress the absence of hydraulic fracturing data for the Liren Tunnel, this study develops an intelligent inversion method integrating the Whale Optimization Algorithm (WOA) with a Backpropagation (BP) neural network. A three-dimensional numerical model was established using FLAC3D, and displacement monitoring data (including crown settlement and horizontal convergence) were used to constrain the boundary stress parameters. Eighteen forward simulations based on the U18^∗(3) uniform design generated training samples, establishing a nonlinear mapping relationship between deformation responses and stress boundary conditions. The WOA was employed to globally optimize the initial weights and biases of the BP neural network, significantly enhancing convergence speed and inversion accuracy. The developed WOA-BP model achieves an average inversion error of 5.73% for vertical deformation, which is 5.48% and 1.61% lower than that of the traditional iterative method and the stan-dalone BP model, respectively. For horizontal deformation, the average inversion error is 6.85%, corresponding to reduction of 6.27% and 4.18%. In addition, the proposed method requires only about one-third of the iterations needed by the BP model. These results indicate that the WOA-BP method offers a highly accurate and computa-tionally efficient solution for estimating in-situ stress fields in soft rock tunnels, providing practical guidance for deformation control and support optimization in similar engineering contexts.

The rock mass integrity is closely related to the stability of the surrounding rock in the tunnel boring machine (TBM) tunnel, and has a crucial impact on the optimization of tunnelling control parameters and support de-cisions. The current research on the perception of surrounding rock information in TBM tunnel based on rock fragment images, only roughly identifies the rock mass grade or customized class, with little attention paid to more detailed rock mass integrity. This research proposes an accurate method for recognizing the rock mass integrity grade in TBM tunnel based on the rock fragment image and Squeeze-and-Excitation (SE)-enhanced Inception-Visual Geometry Group (VGG)19 network. Firstly, the data acquisition systems are developed and the relevant data are collected. Then, the relationship between the particle size distribution of rock fragments and the rock mass integrity is analyzed. Finally, a novel SE-enhanced Inception-VGG19 (SI-VGG19) network is de-signed and a model for recognizing the rock mass integrity grade is established. The ablation experiments show that architecture and class weight optimizations can increase the F₁ value of the SI-VGG19 model on the test set by 2.0% and 1.6%, respectively. The F₁ value of the proposed model on the test set is as high as 0.915, which is 13.1%, 8.3%, 3.9%, 10.2%, 9.0%, 19.7%, 10.9%, and 18.0% higher than that of AlexNet, Xception, VGG19, ResNet50, InceptionV3, MobileNetV2, DenseNet121, and EfficientNetV2B3 models, respectively. Therefore, the proposed method is excellent which can provide guidance for disaster warning and construction optimization of TBM tunnel.

The shear behavior of rock joints with asperity damage in polycrystalline rock is significantly influenced by grain-scale heterogeneity, a critical factor affecting the stability of underground engineering projects. While Peri-dynamics (PD) has been increasingly applied to analyze shear behavior of rock joints, existing models often idealize the rock as an isotropic continuum, neglecting its intrinsic microstructural heterogeneity, elastoplastic response, and detailed joint surface morphology. To address these limitations, this study proposes an elastoplastic fracture model within the PD framework to simulate the shear behavior of polycrystalline rocks. The proposed model features: (1) an explicit representation of polycrystalline structures with diverse mineral compositions, and (2) an accurate characterization of joint surface morphology using joint roughness coefficients (JRC). The model is validated through three-point bending test simulations, which show good agreement with experimental results. Direct shear tests are then simulated to systematically investigate the influences of three key factors: JRC (10.2-17.5), constant normal stiffness (0-0.01 mm), and mineral composition (plagioclase 40-60%, pyroxene 16-24%, biotite 8-12%). The results indicate that while stochastic mineral distribution has a negligible impact on shear strength of rock joints with identical compositions, an increase in plagioclase content enhances strength, and a higher JRC prompts a transition from localized to diffuse damage patterns. These findings provide valuable insights for the design and stability assessment of deep underground engineering where rock joints behavior is a governing factor for overall system performance.