Abstract: In the research of grade of crude ore into concentration plant, it is the critical step of optimizing grade of crude ore to establish the correct mathematic models of subsystems. In ore processing subsystem, the relations among grate of crude ore, rock-mixing ratio and concentration ratio is complicatedly nonlinear. In this paper, the hidden layer structure and learning parameters (eg. learning rate and momentum) of neural network (NN) are optimally determined by the improved genetic arithmetic, then the improved evolutionary NN is used to map the complicated nonlinear relations among grate of crude ore, rock-mixing ratio and concentration ratio. In Jinshandian Iron Mine, the NN model is integrated to the decision support system for optimizing the grade of crude ore into concentration plant, resulting good effect that 162300t iron concentration too many is recovered per year.