In this paper a model for the components of materials based on neural network is introduced. This model is an improved four-layer feed-forward radial basis function (RBF) neural network that uses different training ways according to different characteristics of the patterns, and adds the centers in hide-lay for increasing training speed. The regional mapping from the input's parameters to the output's prevents the network from "excessive learning" and improves the ability of recognition. The simulation of CaO-Al_2O_3-SiO_2 system shows the validity of this method.