Abstract: The short and long-term mechanical properties and deformation characteristics of rocks significantly impact the engineering project’s long-term stability and safety. Traditional constitutive models struggle to uniformly describe the short- and long-term mechanical properties of various rock materials. In contrast, deep learning methods can predict these properties without additional elastic or plastic parameters and constitutive laws. The long short-term memory (LSTM) deep learning algorithm is well-suited for time series data tasks and excels in predicting the short- and long-term mechanical properties of rocks. This study utilized the LSTM algorithm to construct sequence data based on the triaxial compression loading path and stress relaxation over time. A prediction model for the mechanical properties of gray sandstone under conventional triaxial compression and stress relaxation was established. Comparison with experimental data validates the accuracy of the deep learning-based short- and long-term rock mechanical prediction model. To enhance the model’s engineering application value, the LSTM constitutive model was embedded into the finite element (FEM) framework for numerical implementation and applied to gray sandstone simulation. The comparison results show that the LSTM-FEM method better predicts the short- and long-term deformation characteristics of rocks.

Key words: short- and long-term mechanical properties, long short term memory, constitutive relation, finite element method