Abstract: The authors systematically examine the application progress of deep learning in tunnels and underground engineering across seven key areas: parameter inversion analysis, construction machinery parameter prediction and optimization, construction and operation process control and risk assessment, tunnel safety monitoring and defect detection, tunnel structural health prediction, surrounding rock classification, and tunnel face image recognition and classification. The analytical results reveal the following: (1) The theoretical framework for parameter inversion has reached a relatively mature stage. The future of inversion methods in tunnels and underground engineering lies in the development of multivariate intelligent inversion models that integrate advance monitoring technologies, computer algorithms, and simulation techniques to achieve high accuracy and efficiency. (2) The accurate prediction of tunneling parameters is crucial for optimizing the performance of construction machinery and facilitating intelligent tunneling processes. Considering the correlations and differences among tunneling parameters further improves the generalization capabilities of prediction models. (3) The high integration of temporal and spatial multivariate monitoring data with data-driven risk control methods offers intelligent management solutions, facilitating dynamic design and information-based construction during the tunnel construction and operation phases. (4) Computer vision technology, combining feature fusion-based deep neural networks with adaptive pixel level segmentation algorithms, reduces the cost of defect detection, ensuring the applicability of intelligent disaster prevention and safety monitoring systems at engineering sites. (5) By integrating structural health monitoring methods with deep learning algorithms and physical mechanisms, optimized systems improve the accuracy and reliability of tunnel stability and deformation predictions. (6) A deep learning framework based on multivariate information acquisition technologies extracts key parameters of rock mass structural planes and transforms them into quantitative indices, enabling intelligent classification of surrounding rocks under different geological environments and construction methods. (7) Image processing technologies and deep learning algorithms automatically extract valid features from complex tunnel face images for accurate crack feature recognition and quantification. The authors  identify existing challenges for applying deep learning to tunnels and underground engineering, which includes difficulties in real-time data sharing and the lack of evaluation standards for model prediction accuracy. Considering the growing emphasis on intelligent, green, and sustainable construction in tunnel and underground spaces, the authors outline future prospects that involve advancements in deep learning theory and its engineering applications, the development of intelligent disaster prevention technologies for tunnel structures, and new tunnel construction methods, aligning with the goals of the "Carbon Peaking and Neutralization Policy".

Key words: tunnels and underground engineering, deep learning, multivariate monitoring data, lining disease identification, intelligent disaster prevention system