Hybrid physics–data prediction approach for interpretable deformation prediction in pre-support tunnel construction

Deformation prediction during pre-support tunnel construction remains a major challenge: conventional methods lack adaptability, while deep learning models often sacrifice physical interpretability for accuracy. This study addresses the research question of how to achieve predictions that are both accurate and physically consistent under dynamic tunnel construction. We propose a hybrid physics–data prediction approach (HPDPA) that integrates a physical–mathematical model grounded in Winkler elastic foundation beam theory with a physics-informed attention long short-term memory (PI-ALSTM) network. Through a bidirectional collaboration mechanism, the PI-ALSTM supplies data-driven variables to refine the physical–mathematical model, while the latter informs the learning process of the data-driven model, ensuring consistency with physical mechanics and adaptability to the updated site-specific conditions. When validated on real-world pipe-roof tunneling data, HPDPA achieves high predictive precision (MAE = 1.11 mm, RMSE = 0.97 mm, MRE = 0.07) and markedly outperforms purely data-driven baselines. These results demonstrate the framework’s capability to provide physically trustworthy, high-resolution, and real-time guidance for tunnel construction, thereby contributing to the digitalization process of risk management in pre-support tunnel projects.