Abstract: Marine sediment clay has higher creep characteristic. Traditional rheological constitutive models built under each level loading are impractical and hard to describe the nonlinear properties of rheology of rock and soil really. So, the BP neural network is introduced in the text. It has strong nonlinear mapping ability and fault-tolerance ability. Through improving the algorithm of BP, the BP constitutive model is built directly. The model is based on the creep data of laboratory direct shear experiment. It could avoid building a necessary complex constitutive equation in order to meeting the variable law of trial curve and creep character. At the end, the modeling method is tested by the creep experimental data of soft clay of Shanghai and the features of BP neural network rheological model on describing creep curves of soft clay are discussed. It is shown that the rheological model is simple and can effectively describe nonlinear creep of soil clay with better prediction.