Abstract：Study shows that there is normalization characteristic about the stress-strain curves of mediate-dense sands in the range of routine confining pressure. The triaxial test data are normalized by choosing proper normalization parameters. The neural networks are trained by regarding the normalized data as training samples，and then the constitutive model of sand described by neural networks is obtained. The emulation value of the neural networks agrees with well，which shows that the proposed modeling method is reasonable. It can achieve probabilistic optimization automatically based on all test data by using the modeling method，and can reduce the interference of noise signal，and lower the influence caused by dispersive test data.
