Unfrozen water content (UWC) plays a critical role in determining the thermal, hydraulic, and mechanical properties of frozen soils. Existing empirical, semi-empirical, and theoretical models for UWC estimation have limitations in terms of accuracy as well as generalizability. To address these challenges, the present study explored the application of six machine learning techniques to predict UWC in frozen soils: Random Forest (RF), eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), K-Nearest Neighbors (KNN), Support Vector Regression (SVR), and Backpropagation Neural Network (BPNN). Considering the UWC hysteresis phenomenon between the freezing and thawing processes, experimental UWC data collected from the literature were separated into two sub-datasets: freezing branch dataset (FBD) and thawing branch dataset (TBD). Based on that, a comprehensive framework integrating Bayesian optimization and 10-fold crossvalidation was established to optimize the six models' hyperparameters and to evaluate their performance. The results highlighted significant variations in the predictive capability among the six machine learning models, with ensemble methods (i.e., RF, XGBoost, LightGBM) generally demonstrating superior accuracy. Feature importance analysis, robustness checks, and uncertainty quantification further elucidated the strengths and limitations of each model. The present study provides profound insights into the selection and application of machine learning models for accurately modeling the properties of frozen soils for cold regions science and engineering.

In order to study the impact of surface roughness on the cyclic shear characteristics of the Soil-Rock Mixture and concrete interface, a series of cyclic shear tests were conducted using a large indoor direct shear apparatus. The effects of three concrete surface roughness coefficients JRC (0.4, 9.5, 16.7), five rock content levels (0%, 25%, 50%, 75%, 100%), and three cyclic shear displacement amplitudes (1, 3, 6 mm) on interface cyclic shear stress and Soil-Rock Mixture shear deformation were analyzed. A Bidirectional Long Short-Term Memory (BoBiLSTM) model was proposed, utilizing Bayesian optimization and k-fold cross-validation for hyperparameter tuning to streamline the model parameter selection process and enhance the prediction accuracy of the stress-strain relationship under cyclic loading. The experimental results show that, under five rock content levels, as the concrete surface roughness coefficient and cyclic shear displacement amplitude increase, the interface average peak shear stress increases accordingly. The interface average peak shear stress of the sample with 75% rock content is the highest; in terms of vertical displacement, the sample with 50% rock content has the maximum displacement, while the sample with 25% rock content has the minimum. The two types of samples show different soil deformation patterns in the two shear directions during the cyclic shearing process; as the shear displacement amplitude increases from 1 mm to 3 mm and 6 mm, the greater the concrete surface roughness, the smaller the change in shear stiffness and damping ratio. Compared to traditional Long Short-Term Memory (LSTM) models, the BoBiLSTM model demonstrated improvements in the average metrics of R2, RMSE, and MAPE by 0.32%, 57.25%, and 72.32%, respectively.