A new approach for the prediction of azeotrope formation between components in a mixture, that does not require vapor-liquid equilibrium calculations, is presented. The method employs neural networks to correlate azeotropic data for binary mixtures with a series of macroscopic and microscopic properties of the pure components, without explicit consideration of non-ideality of mixture. The model fails to make a clear prediction regarding azeotropy in only a relatively small number of situations in which structurally homologous molecules are known to exhibit quite distinct azeotropic behavior. (C) 2003 Elsevier Ltd. All rights reserved.