Abstract: A simulated-annealing-based immune algorithm (SAIA) is presented in the paper. By imitating the biological immune system's characteristics of immune recognition and learning to respond to invading antigens, the algorithm can restrain the degenerate phenomenon by using the immune mechanism based on the simulated annealing to maintain the individual diversity. In a neural network prediction model of the silicon content in hot metal, the SAIA is used to optimize the connection weights of a multi-layer feed forward neural network to improve the prediction precision. The testing and simulating results are satisfactory.