A data and physics driven method based on the bidirectional and nested model (BNM) was proposed to predict the long-term settlement of shield tunnels during operation. Unlike the physics-informed neural network (PINN) method, where physical knowledge was directly embedded into the neural networks, the BNM integrated machine learning with analytical and numerical simulations, allowing physical laws to be internally represented by the nested analysis or numerical submodels, thereby enhancing physical coupling capability. Taking the Nanjing East Road—Lujiazui section of Shanghai Metro Line 2 as a case study, a BNM-based long-term settlement prediction model was developed and compared with purely data-driven approaches. The results indicate that single machine learning models are sensitive to data partitioning and perform inconsistently across validation sets, showing limited engineering applicability. In contrast, the proposed BNM model achieves stable and high prediction accuracy under multiple and complex influencing conditions. Combining the advantages of data-driven method and physics-based method, the BNM is suitable for long-term settlement prediction and engineering decision support for shield tunnels under complex operational environments.