The estimation of kinematic viscosity of petroleum crude oils and fractions with a neural net

Abstract This paper illustrates how a neural net, a three-layered perceptron, can be trained to estimate viscosities for undefined crude oils and fractions. Three Saudi-Arabian crude oils were employed to illustrate the use of the neural net to approximate the relation in a very simple manner with no need for a priori knowledge of the system. This empirical correlation was accurate to 98.74% if tested on experimental data not used during training, which is a fivefold improvement on average results obtained by two recently-proposed equations to estimate the viscosity of hydrocarbons. Although the neural net equation seems to be less transparent than former correlations, a method called backward analysis is proposed to analyze the weight matrix of the neural net in order to gain valuable insight into the viscosity system.