Abstract: There is a great difference in effects caused by uncertainties in the actual foundation water load and in different ways. If the foundation water load is put artificially as a surface load or as a stable seepage body load through numerical calculation in the optimization back analysis， the parameters obtained from inversion are debatable. In this paper, a relative displacement of monitoring stations is taken as an input， and the dam concrete， batholith material parameters and the water head of the face of the dam foundation at a certain depth point are used as an output， thus the identification neural network model for the uncertainty of foundation water load is established. By adopting the uniform design principle of material parameters combination， the saturated foundation non-stationary seepage analysis is made to get different seepage body load distributions， and samples to learn， so as to train a good network in describing the nonlinear relationships of the concrete dam， material parameters， ground water load and dam deformation. The water pressure component separated out from the measured displacements is put into the trained network， which can automatically recognize the dam concrete and batholith material parameters as well as the foundation water load. The calculation examples show that the establishment of a neural network model for the identification of the uncertainty of the foundation water load is feasible.