Predicting soil stress–strain behaviour with bidirectional long short-term memory networks

Purpose Artificial intelligence, particularly deep learning (DL), has increasingly influenced various scientific fields, including soil mechanics. This paper aims to present a novel DL application of long short-term memory (LSTM) networks for predicting soil behaviour during constant rate of strain (CRS) tests. Design/methodology/approach LSTMs are adept at capturing long-term dependencies in sequential data, making them suitable for predicting the complex, nonlinear stress–strain behaviour of soil. This paper evaluates various LSTM configurations, optimising parameters such as step size, batch size, data sampling rate and training subset size to balance prediction accuracy and computational efficiency. The study uses a comprehensive data set from numerical finite element method simulations conducted with PLAXIS 2D and laboratory CRS tests. Findings The proposed LSTM model, trained on data at lower stress levels, accurately forecasts soil behaviour at higher stress levels. The optimal LSTM setup achieved a median error of 3.59% and 5.10% for numerical data and 3.86% for laboratory data, presenting the setup’s effectiveness. Originality/value This approach reduces the required time to complete extensive laboratory testing, aligning with sustainable industrial practices. The findings suggest that LSTM networks can enhance geotechnical engineering applications by efficiently predicting soil behaviour.