This paper reports on the chemometric identification of the 3 butter types (cultured cream-, sweet cream- and mildly soured butter) by use of neural networks which is suitable to minimize the analytical expenditure. The results of compositional analyses of butter samples were used as inputs for a three layer feed-forward back-propagation network. The data were randomly divided in three sets for training, testing and validation. The network A with two inputs (pH-value, citric acid) and two hidden nodes was sufficient to give correct results. Thereby, the differentiation of butter types is considerably minimized as compared with previous approaches, e. g. the calculation of butter type indices.