Modeling of frozen soil-structure interface shear behavior by supervised deep learning

This paper proposes a data-driven approach to characterize the interface shear behavior between frozen soil and structure surface, which can be regularly encountered in emergency repairs on tunnels using artificial ground freezing. Experimental data of 32 constant normal stress shear tests was compiled and used for supervised training. Exposed temperature, soil moisture content, and normal pressure were used as predictors and shear stress at a specific relative displacement as output. The proposed framework, integrating either backpropagation neural network (BPNN) or bidirectional long short-term memory (Bi-LSTM), was assessed by k-fold cross-validation without or with gaps on a limited data sample. Three performance indicators (RMSE, MAPE, and R2) were employed, showing that both BPNN and Bi-LSTM based computing models can reliably reproduce the peak and residual adfreeze strengths and interface softening behavior as ice bond breaks. Bi-LSTMs generally outperform BPNNs as the former are better-suited for analyzing a sequence of discrete-time data. The feasibility of using supervised deep learning in interface shear behavior modeling is demonstrated. The rapid advancement of digital technologies offers an opportunity for practitioners in artificial ground freezing and other construction projects to make informed decisions using such data-driven methods.