Development of an artificial neural network for the prediction of relative viscosity of ethylene glycol based nanofluids

In this paper, we have developed an artificial neural network (ANN) model for the prediction of the viscosity of ethylene glycol-based nanofluids using data available in the literature. To develop the model, 377 data points were taken from the available literature. The data includes MgO, Y3Al5O12, In2O3, Ag, SiO2, Fe, Mg(OH)(2), ZnO, SiC, Al2O3, CeO2 and Ce3O4 nanoparticles. The inputs given to the ANN model were the diameter of the nanoparticles, temperature, and concentration of the nanoparticles, whereas output was the ratio of dynamic viscosity of the nanofluids to that of the base fluid. The ANN model was trained using 80% of the dataset and the rest of the dataset was used for testing the performance of the developed model. In order to prevent the model from getting overfit, dropout layers were also used. The trial and error method was used to find the optimum model. The optimum model consisted of 2 hidden layers and 45 neurons in both the hidden layers. The developed model shows good performance with the value of mean square error for the training data and test data being 3.9E-04 and 4.4E-04, respectively. The value of correlation coefficient (R) for the training data and test data was found to be 0.9962 and 0.996, respectively. Despite the high number of neurons in hidden layers, performance parameters reveal that there is no overfitting in the model. A comparison between the experimental values and the values predicted by the ANN model is also done in this paper.