Modeling axial strain-deviatoric stress response of frozen sands with enhanced LSTM approach

Abstract

Data-driven approaches hold promise for modeling the highly nonlinear stress-strain behavior of frozen soils in comparison to standalone machine learning models which often overfit and lack generalizability. To address these issues, in this study an enhanced fundamental Long Short-Term Memory (LSTM) model with an iterative EXtreme Gradient Boosting (XGBoost) algorithm is proposed to predict the axial strain-deviatoric stress relationship of frozen sands. Using degree of saturation, mean particle size, test rate, initial void ratio, temperature, confining pressure, and axial strain as input parameters, the Local Interpretable Model-agnostic Explanations (LIME) feature-importance analysis identified initial void ratio as the most influential parameter. The adaptability of the fundamental LSTM shows better accuracy than the Random Forest (RF) model and the Multilayer Perception (MLP) model. To reduce discrepancies between predicted and measured results in the fundamental LSTM model, an uncertainty factor was introduced to improve residual accuracy, thereby facilitating the development of an XGBoost-optimized LSTM (LSTMXGBoost) model. The resulting LSTMXGBoost model demonstrated strong predictive performance for axial strain-deviatoric stress relationship across a range of triaxial shear test conditions. Analysis of results suggest that there is a good comparison between two key parameters; namely, modulus of elasticity and peak stress that were extracted from a separate triaxial test dataset not used in model training or testing. The results of this study are promising for constructing high-dimensional constitutive models that can be used in numerical simulations of frozen sands behavior for use in geotechnical engineering practice applications alleviating time consuming, cumbersome and expensive experimental test techniques.