Abstract: This paper tries to explain the reason that the result obtained from solving the inversion of interfaces using BP neural networks is not satisfactory.Firstly, the concept that capacity of the training sample set is used for BP neural network was introduced in terms of information quantity in this paper, and its definition was also given, The principle and method used in forming a training sample set were presented.Secondly, the basic procedure solving the inversion of interfaces using BP neural networks was given.Finally, the above-mentioned concept and method were applied to an examlpe of inversion of interfaces.By analysing the example carefully, the author pointed out what difficulty was encounte-tred in solving the inversion of interfaces using BP neural network.