Prediction of shear strain and excess pore water pressure response in liquefiable sands under cyclic loading using deep learning model

抄録

In this study, an attempt is made to predict the shear strain and excess pore water pressure ratio (r_u) response of liquefiable sands using deep learning (DL) models. The DL model, such as long short-term memory (LSTM), is considered for predicting the response. The inputs of the models are basic soil properties, including applied shear stress time history, existing initial vertical effective stress, relative density, coefficient of uniformity, and mean grain size of the soil. The databases considered for the DL model training and testing are the cyclic laboratory test data of Ottawa F-65 and Nevada sands. For r_u model training, suitable inputs are considered based on the available domain knowledge on the generation of excess pore water pressure during liquefaction in sands. Shear strain is also an essential input parameter to predict the excess pore water pressure ratio during liquefaction. Therefore, an additional shear strain model is developed to predict the shear strain time history for the testing datasets using the same training datasets. Then this shear strain model is used along with the r_u model to predict the excess pore water pressure response for separate testing datasets. The predicted responses of shear strain and excess pore water pressure ratio agree well with the actual responses for testing datasets and perform well in terms of the evaluation metric. However, the models are trained using the limited datasets of specific soil types and their performance has to be tested for various global soil types under different loading conditions (e.g., transient loading). The study concluded with recommendations for improving the DL models for the cyclic response of other sands for seismic stability assessments.