Mechanical parameters inversion method based on multi-layer GRU architecture and Bayesian optimization considering time-dependent deformation of tunnel

ABSTRACT: Accurate acquisition of the mechanical parameters of surrounding rock is fundamental for the assessment of tunnel stability. This paper introduces a surrogate model and inversion framework for tunnels based on deep learning and Bayesian optimization algorithms, which considers the creep characteristics of surrounding rock and burial depth variations among different tunnel segments, while making full use of displacement data from monitoring points at each tunnel section. Initially, a constitutive model of creep based on internal variable thermodynamics is adopted to simulate the time-dependent evolution process of surrounding rock deformation. Subsequently, the Latin hypercube sampling method is utilized to construct parameter samples, and numerical simulations are employed to build a training dataset. Following this, hyperparameter optimization of deep learning models based on multi-GRU architecture is conducted to establish complex mapping relationships between parameters and time-dependent deformations. Finally, measured deformations and burial depths are used in conjunction with Bayesian optimization algorithms for inversion. The proposed inversion framework was applied to the JLL tunnel, achieving rapid sectional inversion of the mechanical parameters for targeted tunnel segments along the entire line. The results demonstrate an average inversion accuracy of 93.9%. This inversion approach can provide support for tunnel support design and stability calculations.