In this paper, we proposed a SOM neural network which based on kernel function. It adopts kernel function to replace Euclidean distance, and take it as one criterion to estimate the matching degree between the input pattern and the connection weight. It accumulates knowledge by the process of learning to the input pattern and connection weight adjustment. It is unsupervised learning, it has the ability of self-learning, self-adaptive and self-stability. It has shown fascinating characteristic when being used in silicon content prediction of molten iron.