A digital twin system for long-term slope deformation prediction based on an adaptive updating surrogate model

Predicting long-term rheological deformation of slopes under complex geological conditions is a major challenge in geotechnical engineering. Digital twin technologies have emerged as an effective solution by integrating numerical models with real-time monitoring for enhanced predictive performance. A key component of digital twins is data assimilation, which enables dynamic parameter updating using observational data. However, repeated calls of high-fidelity numerical models during data assimilation impose a significant computational burden. To address this issue, this study proposes a digital twin-based prediction framework that integrates an adaptive surrogate model with the Markov Chain Monte Carlo (MCMC) method for efficient Bayesian updating of rheological parameters. Unlike conventional surrogate models, the proposed adaptive strategy continuously updates the surrogate model in the high posterior density region, enabling real-time calibration and improved alignment with physical behavior. By integrating both continuous monitoring data and high-fidelity simulation results, the framework better embodies the core philosophy of digital twins and progressively approximates real-world conditions. The framework is validated through synthetic cases and applied to the right-bank slope of the Baihetan hydropower station. Results show that elastic parameters converge rapidly due to strong geological constraints, while viscous parameters remain uncertain in structurally complex zones. The adaptive surrogate model successfully captures both initial rapid deformation and long-term creep behavior, closely matching field measurements. This study demonstrates the potential of digital twins with adaptive learning capabilities for reliable and efficient slope deformation prediction in large-scale geotechnical systems.