Integrated Machine Learning Using Autoencoder and Multilayer Perceptron Approaches to Predict Stress–Strain Curves of Frozen Soil–Rock Mixtures

Machine learning is extensively used in constitutive modeling of geomaterials and is critical for enhancing computational analyses in geotechnical engineering. In the area of geomechanical modeling, the primary application of machine learning techniques is to use discrete data points for model training and testing. However, this methodology frequently fails to accurately capture the stress–strain behavior of geomaterials because of imbalanced training data sets, which undermines model performance and accuracy. In addition, the limited interpretability of these models hampers their practical application and understanding. To address these shortcomings, a machine learning framework specifically developed for the analysis of frozen soil–rock mixtures is introduced. The proposed model uses an autoencoder to extract features directly from individual stress–strain curves. The extracted features serve as training outputs for the multilayer perceptron, and the test conditions are employed as inputs to train the model. The decoding function of the autoencoder is employed to reconstruct the stress–strain data. The average fit between the model predictions and the actual results exceeds 93%. Shapley additive explanations are further incorporated to quantify the impact of the test conditions, revealing that temperature exerts the most substantial influence, followed by block proportion and confining pressure. Notably, an analysis reveals that the proportion of block significantly affects postpeak stress–strain responses, highlighting its comparable importance to temperature. This study identifies key test conditions influencing the mechanical behavior of frozen soil–rock mixtures, providing geotechnical engineers and researchers with a reference tool for predicting their properties.