Thermal cracking of hydrocarbons converts them into valuable materials in the petrochemical industries. Multiplicity of the reaction routes and complexity of the mathematical approach has led us use a kind of black-box modeling-artificial neural networks. Reactor feed type plays an essential role on the product qualities. Feed type is a qualitative character. In this paper, a method is presented to introduce a range of petroleum fractions to the neural network. To introduce petroleum cuts with final boiling points of 865 F maximum to the neural network, a real component substitute mixture is made from the original mixture. Such substitute mixture is fully defined, it has a chemical character, and physical properties can be simply retrieved from databases. The mixture compositions are defined with the aid of an optimization algorithm-interval method. The obtained TBP curves of substitute mixture are in good agreement with the experimentally obtained curves. Nine single carbon structural increments will be the representative of 93 real component compositions in order to make the topology of the neural network smaller and hence to have a less complex model.