Prediction of The Stress-Strain Curve of Cement Admixed Clay by using LSTM Recurrent Neural Network

Abstract

This study used the recurrent type of neural network LSTM to predict the stress strain of cement mixed clay in the triaxial test. The features for simulation were mixing ratio water and cement, mean stress, deviator stress and vertical strain. The best architecture for the neural network was proposed in this study with the lowest error. The LSTM was the best model with the lowest error among other types of recurrent neural network, GRU and SimpleRNN. LSTM with 2 time steps was the best architecture to predict the stress strain characteristic of clay mixed. The prediction model can simulate the stress-strain relationship with an average absolute error of 4 %.