Accurately forecasting tunnel convergence is essential for safety management and design optimization during excavation. Traditional numerical models often struggle with heterogeneous ground conditions, incomplete field monitoring data (FMD), and uncertainty in input parameters. To address these limitations, this review examines recent developments in artificial intelligence (AI) for tunnel-deformation prediction, with emphasis on methods that improve transparency, interpretability, adaptability, and reliability. We analyze more than fifty peer-reviewed studies published between 2018 and 2024, covering three major categories of AI techniques: machine-learning (ML), hybrid metaheuristic optimization, and deep-learning (DL). A structured review methodology is applied to identify common data sources, preprocessing steps, model architecture, and validation practices used in tunnel engineering research, with brief consideration of methodological consistency and potential limitations reported in the literature. Based on this synthesis, we propose an interpretable-to-deep (I2D) framework consisting of five layers: field monitoring data, data processing and feature engineering, hybrid AI, deep sequence learning, and prediction and decision support. Rather than focusing on individual algorithms, the framework highlights how different levels of AI con- tribute complementary strengths, including interpretable ML techniques that provide insight into feature importance and model behavior, optimization for stable learning under limited data, and deep temporal modelling for long-term forecasting. Validation strategies typically assess accuracy using standard statistical measures and quantify uncertainty through prediction intervals. Increasingly, recent studies include interpretability techniques to support consistency between model outputs and expected geotechnical behavior. The overall findings indicate a clear shift from standalone predictive models toward integrated, data-centric systems, with improved robustness, interpretability, and potential for real-time decision support. This review provides a practical roadmap for future research on AI-assisted tunnel deformation prediction and supports the development of transparent, reliable, and scalable decision tools for underground engineering practice.
Saadati et al. (Wed,) studied this question.