Prediction of surface tension of alcohol + water solutions using Feed-Forward neural networks and thermodynamic model at various temperatures

Feed-Forward Neural Networks(FFNNs) are among the most important neural networks that can be applied to wide range of forecasting problems with a high degree of accuracy[1-2]. The aim of this study, was predicting surface tension of aqueous solution of methanol, ethanol, 1-propanol and 2-propanol in range temperatures of 293.15-323.15K using with two entrance variables, mole fraction and temperature, one hidden layer(with 1-10 neuron) and one output neuron. In this work, a set of experimental data of alcohol/water mixtures at various temperatures have been collected from the literature [3].The optimal FFNN structure model was determined based on the maximum value of R2 and the minimum value of the APD of the testing set for each system. The obtained results for R2 and APD values, showed that the best FFNN architectures for binary mixtures of water/methanol, ethanol, 1-propanol and 2-propanol are (2:4:1), (2:5:1), (2:3:1) and (2:6:1), respectively. As it is obvious from Fig. 1, the calculated values of training, validation and testing sets are located around the bisection, and this indicates the accuracy of the results and the ability of the used FFNN model for predicting the desired property (water/methanol systems). In Fig. 2, the differences between the experimental and calculated values for water/ 1-propanol system are potted versus the experimental data. This figure shows that the range of error for surface tension is (-0.38 to 0.33), other systems had a same pattern. The obtained results confirm that the FFNN is a powerful method for predicting the surface tension of high-polar binary mixtures of water/alcohols. Finally, to check the performance of the FFNN model, its estimations are compared with thermodynamic model base on chemical equilibrium and discussed from a theoretical point of view [4].