Multi-fidelity model for real-time prediction and reconstruction of 3D greenfield ground response to tunnelling with quantified uncertainty

Accurately real-time predicting and reconstructing three-dimensional ground deformation are essential for ensuring the safe progression of shield tunnelling and the stability of the surrounding environment. However, conventional machine learning, analytical, and finite element methods struggle to predict ground deformation effectively. The paper proposes a multi-fidelity model for real-time prediction and reconstruction of three-dimensional greenfield ground response to tunnelling with uncertainty quantified. In the model, an analytical solution (AS) provides low-fidelity deformation predictions, ensuring a physics-consistent global distribution and deformation evolution trend. Besides, a Random Forest-based high-fidelity residual model (RFRM) is designed to refine the low-fidelity model, and the infinitesimal jackknife is utilised to quantify uncertainty. AS and RFRM are integrated to develop the multi-fidelity prediction model (AS-RFRM). Four case studies were conducted to validate the AS-RFRM, and the comparative analysis results indicate that the AS-RFRM achieves a 4%–60% improvement in predictive performance compared with alternative methods. The AS-RFRM is capable of forward-predicting deformation induced by the next three consecutive tunnel rings at least, and its quantified uncertainty intervals generally capture the trend of settlement development. AS-RFRM offers robust forward predictions with sparse data, facilitating timely construction interventions and reducing monitoring costs.