This paper is concerned with using the neural network to predict the model for the materials composition. A four layer feed forward radial basis function neural network is built, so that a nonlinear mapping through changing the parameter of the nonlinear activity function is realized, which reflects on the linearization when adjusting the connecting weight. Therefore, it can accelerate the learning speed and reduce the quantity of calculation, also the problem of local minimum and the slow rate of convergence of the back propagation neural netwok are avoided. In this paper CaO Al 2O 3 SiO 2 system is discussed as an example. Simulation shows that neural network can be used to predict the materials composition successfully.