The application of digital twin technology in shield tunnelling can significantly enhance the intelligence level of tunnel excavation while reducing uncertainty and risks. Nevertheless, the core lies in how to achieve real-time prediction, analysis, and feedback. To this end, we propose a physics–data dual-driven real-time framework based on a finite-element surrogate modeling approach[1-3]. This framework integrates high-fidelity finite element numerical models with long short-term memory (LSTM) neural networks, thereby combining the reliability of physics-based models with the computational efficiency and real-time performance of data-driven surrogate models. Furthermore, the proposed framework demonstrates strong performance in small-sample cases, making it particularly suitable for practical engineering applications where data availability is limited.The framework consists of a three-step architecture. Firstly, the offline surrogate model training is provided, in which large-scale finite element simulations are performed to generate training datasets, establishing mappings from geotechnical and construction parameters, to the shield tunnelling status. These simulation data are then used to train neural-network-based surrogate models, enabling efficient reactions of the complex nonlinear relationships between system inputs and tunnelling responses. The second step is dynamic updating of the numerical model, in which real-time monitoring data acquired during the construction process are assimilated to rapidly invert geotechnical parameters through the surrogate model, thereby enabling continuous updating of the numerical model to maintain consistency with the actual tunnelling state. Thirdly, online physics–data dual-driven model is running. Taken machine status as optimization objectives, operational parameters are rapidly optimized through the surrogate model. The optimized parameters are input into a high-fidelity numerical model to predict the future status, enabling forward-looking assessment of the excavation. The proposed method has been applied to a cross-sea tunnel digital twin project in South China, where it provides accurate predictive performance and real-time decision support, significantly shortening the construction time and enhancing construction safety.