Abstract: Deformation serves as a pivotal indicator for the stability assessment of geotechnical structures, as exemplified by the deflection of deep excavation retaining systems and the convergence of shield tunnel segments. Addressing the limitation that conventional deformation prediction methodologies are predominantly data-driven—heavily contingent upon the quantity and fidelity of available data—we propose a physics-data dual-driven prediction framework anchored in multi-fidelity networks. The framework comprises two distinct modules. Initially, a low-fidelity sub-network is constructed utilizing the deep operator network (DeepONet) architecture, designed to mine fundamental deformation characteristics inherent in numerical simulation results and extract the physical modes of structural deformation. Subsequently, a residual-based high-fidelity sub-network is established, wherein high- and low-fidelity data are explicitly correlated at the input and output interfaces. Employing a long short-term memory (LSTM) architecture, this module realizes a physics-guided data-driven approach, whereby the low-fidelity solution provides guidance and correction to the high-fidelity prediction. By embedding physical paradigms, the framework effectively constrains the training of the data model, thereby mitigating the reliance on purely empirical data and surmounting the limitations imposed by data quality. Finally, the proposed framework is applied to the Taihu Tunnel deep excavation project. The results demonstrate that the predictive accuracy of the physics-embedded dual-driven methodology is significantly superior to that of a singular data-driven approach. For the two prediction tasks, the mean scores are 0.93 and 0.96, representing improvements of 0.37 and 0.035, respectively, relative to the single data-driven method; the mean MAEs are recorded at 0.265 and 0.21, corresponding to reductions of 28% and 27%; and the mean MSEs are 0.125 and 0.083, indicating substantial reductions of 84% and 47%.

Key words: geotechnical structures, deep foundation pit, deformation prediction, physics-data fusion-driven, multi-fidelity methods