Steam distillation process improves oil recovery processes involving steam injection up to 50%. Due to its immense effect on oil recovery, several attempts have been made to simulate this process experimentally and theoretically. Since detailed crude oil data is rarely available, a model should be presented to predict the distillate rate with minimum entry parameters. For this purpose, a Multi-Layer Perceptron (MLP) network is used in this research as a new and effective method to simulate the distillate recoveries of 16 sets of crude oil data obtained from literature. API, viscosity, characterization factor and steam distillation factor are input parameters of the network while distillate yield is the result of the model. Thirteen sets of data were used for training the network and three remaining sets were used to test the model. Comparison between the developed MLP model, Equation of State (EOS)-based method and Holland-Welch correlations indicates that the errors of the MLP model for training and test data sets are significantly lower than that of those methods. Also, the MLP network does not require oil characterization, which is a necessary and rigorous step in EOS and Holland-Welch methods. (C) 2009 Published by Elsevier B.V. on behalf of The Institution of Chemical Engineers.