Checking the performance of feed-forward and cascade artificial neural networks for modeling the surface tension of binary hydrocarbon mixtures

To estimate the surface tension of liquid hydrocarbon mixtures as an essential thermophysical property, artificial neural networks (AANs) are used. To develop the AAN model, 25 binary mixtures containing 560 data points in a wide range of temperatures (287.81–343.15 K) and at atmospheric pressure were considered. The performances of two neural networks including feed-forward (FFNN) and cascade neural network (CNN) are compared with different input variables. For both cases, the Levenberg–Marquardt optimization method is used to optimize the weights and biases of the proposed structures with tansig and pureline transfer functions in the hidden and output layers, respectively. It was found that the CNN with the structure of 5-9-1 and input variables of temperature (T), mole fraction (x), molecular weights of both compounds (MW), and mixture critical temperature (Tc-mix) is the optimum model with the average absolute relative deviation (AARD%) = 1.33 and correlation coefficient (R2) = 0.992. The most important feature of the proposed model is its ability to differentiate between isomers and correctly predict their binary surface tensions.