Accurately predicting the permanent deformation of subgrade soils under combined loading and seasonal freeze-thaw (F-T) effects is crucial for optimizing pavement design and maintenance planning in cold regions. However, existing empirical models have limitations in capturing the non-linear deformation behavior influenced by interacting factors. This study developed a deep learning framework, specifically a convolutional neural network-bidirectional long short-term memory (CNN-BiLSTM), for time-series prediction of subgrade strains. Laboratory tests established a comprehensive database to examine soil response across various conditions. Regression analysis revealed the constraints of traditional models. The optimized hybrid architecture directly learned patterns from experimental data, outperforming regression by 29−71 % based on error metrics. Sensitivity analysis confirmed that the model identified primary governing inputs consistent with theory. Hyperparameter tuning with the Subtraction-Average-Based Optimizer further enhanced performance. This data-driven methodology demonstrated the potential of advanced machine learning in simulating complex subsurface deformation behavior under compounding factors in seasonal frozen environments.
Read full abstract