Abstract:Neural network has been nicely applied in pattern recognition and hydrocarbon prediction in applied geophysics. The importance of forward neural network lies in showing the implicit functional relation among known examples, but its generalized application needs to be studied further. Hence we point out the importance of the problem, and put forward the technique for improving the generalization of the neucal network. The technique can reduce overall error and smooth "implicit function" as much as possible in the course of learning, the smoothing degree of the implicit function being indicated by curvature. Computation example shows that our algorithm can obviously improve the generalization of the neural network. The real example is given of how the technique was applied in hydrocarbon prediction in Laohe Oil Field.