Enhancing ground settlement prediction in deep excavations in soft soils through Bayesian updating of time-dependent model biases

Accurate prediction of ground settlements in deep excavations within soft soils remains a critical challenge in geotechnical engineering, primarily due to the complex time-dependent behavior of soils and systematic discrepancies between theoretical predictions and field observations. This study introduces a Bayesian updating framework that integrates Markov Chain Monte Carlo (MCMC) sampling with a bidirectional Long Short-Term Memory (Bi-LSTM) surrogate model to address uncertainties in soil parameters and time-dependent model biases. Validation through a 30.2-m-deep excavation case in Hangzhou demonstrates a robust linear correlation between model bias and normalized excavation time (R 2 > 0.80 in 93.4% of simulations), confirming the time-dependent evolution of model biases. The proposed method achieves remarkable improvements, reducing prediction errors by 85%–99% for short-term forecasts (next three stages) and 17%–93% for long-term predictions, as further validated by the Taipei National Enterprise Center project. Progressive Bayesian updates progressively tighten the posterior distributions of soil parameters and bias coefficients, improving prediction reliability as excavation advances. This work advances geotechnical modeling by systematically addressing the interplay of parameter uncertainty and time-dependent biases, offering a scalable and practical solution for complex, multi-stage excavations in soft soils. The framework’s adaptability to real-time data makes it particularly valuable for adaptive construction management, enabling engineers to optimize resource allocation and mitigate risks associated with ground settlements.