Prediction of Frozen Clay Strength Under Different Temperature Conditions Using Machine Learning Approaches

Frost heave and thaw settlement are among the most widespread and destructive geotechnical hazards in cold regions, posing serious threats to the safety and long-term performance of infrastructure. The initiation and evolution of these hazards are highly dependent on the mechanical properties of frozen soils, such as compressive strength, cohesion, internal friction angle, and deformation modulus. These properties are jointly controlled by temperature, ice content, water content, and freeze–thaw cycling, resulting in strong nonlinearity, temporal variability, and spatial heterogeneity. As a result, conventional laboratory testing and empirical approaches often suffer from high cost, low efficiency, and limited applicability in parameter determination and prediction.　In recent years, machine learning techniques have been increasingly applied to predict soil mechanical parameters due to their ability to handle multi-source data and capture complex nonlinear relationships. However, the strong temperature sensitivity of frozen soil behavior makes it difficult to achieve high prediction accuracy by solely establishing mappings between temperature–moisture–structural characteristics and mechanical responses. This challenge highlights the necessity of data-driven modeling frameworks that explicitly consider stress states and thermomechanical coupling effects.　In this study, a machine learning–based framework was developed to predict the strength characteristics of frozen clay. A total of 116 sets of directional shear test data were used to train and validate four machine learning algorithms. The intermediate principal stress coefficient, principal stress axis orientation angle, mean principal stress, and temperature were selected as input variables, while frozen clay strength was taken as the output. Model performance was systematically evaluated using cross-validation and further verified through comparison with supplementary experimental data.　Based on the optimal model, the distribution of frozen clay strength within a multi-dimensional input parameter space was analyzed. In addition, model interpretability techniques were employed to conduct sensitivity analysis, enabling quantitative evaluation of the relative importance of different input parameters. The results demonstrate that machine learning approaches can accurately reproduce the stress–strain behavior and failure strength of frozen clay, while effectively capturing the complex nonlinear relationships between strength and controlling factors.　Overall, this study shows that machine learning provides a robust and efficient alternative for predicting frozen soil mechanical parameters. The proposed framework enhances prediction.