Artificial Intelligence for Tunnel Seismic Response: A Comprehensive Review

Emerging artificial intelligence and machine learning (AI-ML) techniques are transforming seismic response analysis of tunnel structures by addressing limitations of traditional finite element methods. This comprehensive review examines developments from 2020 to 2025, analyzing validated research from high-impact journals. Modern neural networks—especially convolutional neural networks (CNNs) and long short-term memory (LSTM) models—now predict tunnel damage with over 91% accuracy, compared to the 78–85% accuracy achieved by conventional techniques. These AI systems also run in hours rather than days, making real-time or near-real-time assessment possible. Physics-informed neural networks blend core geotechnical principles with data-driven learning to improve prediction reliability. By fusing data from accelerometers, geological surveys, and live monitoring instruments, multi-source AI models detect damage with more than 90% precision. Machine learning-driven feature importance analysis reveals tunnel depth and soil properties as critical parameters influencing seismic response. Graph neural networks effectively model complex spatial relationships in tunnel networks, while transfer learning enables adaptation across different geological contexts. Inconsistent data formats, the requirement for transparent “explainable AI” to comply with regulations, and smooth integration into pre-existing engineering workflows are still obstacles in the way of these advancements. Future research directions emphasize federated learning for collaborative model development, explainable AI for enhanced interpretability, and physics-constrained algorithms ensuring geotechnical consistency. Looking ahead, federated learning for collaborative model training, explainable AI tools, and physics-constrained algorithms promise to solidify the role of AI in a safer, more resilient tunnel design and monitoring.