Cyclic shear stress-strain prediction of saturated sand based on the unrolled seq2seq model and scheduled sampling

The simulation of stress-strain response of saturated sand under numerous cyclic loadings is of significant interest to the development of constitutive model, and the simulation results form basic data for soil dynamic characteristic estimation, such as pore water pressure build-up, liquefaction strength, dissipated energy, and damping ratio. Traditional constitutive models are usually constructed based on complicated theories and assumptions, and numerous model parameters complicate the calibration process, which limits their application in the rapid evaluation of soil dynamic properties. Although the data-driven model can effectively avoid these problems, it's difficult for a conventional data-driven model to simulate under numerous cyclic loadings due to the simpleness of the model and training strategy. In this study, a novel data-driven constitutive model is proposed to well simulate the stress-strain behaviour of saturated sand and rapidly estimate soil properties only based on a small amount of experimental data. The unrolled sequence-to-sequence recurrent neural network (RNN) is first introduced in soil data-driven constitutive model to modify conventional one-step ahead prediction to multistep one, and a novel RNN training strategy named scheduled sampling, is introduced to reduce the discrepancy of the model behaviour in the training phase and that in testing due to the unknown stress-strain information at next step. The performance of the proposed model is verified through a dataset collected from a triaxial test along with the comparison with feedback neural network, one-step RNN, and non-data-driven constitutive model. Based on the simulation of stress-strain, the dynamic properties of saturated sand are estimated. Both the simulation and estimation results show that the proposed method performs reasonably well.