A novel approach for predicting multiple key properties of water-based nanofluids using artificial neural networks

This study represents the first implementation of a single neural network to forecast multiple fundamental properties of water-based nanofluids rather than employing distinct neural networks for individual nanofluids and properties. For each property, 701 experimental data points for 22 different ([Figure presented], [Figure presented], [Figure presented], ND, [Figure presented], [Figure presented], [Figure presented], [Figure presented], [Figure presented], [Figure presented], [Figure presented], [Figure presented], [Figure presented] in five different mixing ratios, [Figure presented], [Figure presented], [Figure presented], [Figure presented], and [Figure presented]) water-based nanofluids collected from several studies in the literature having particle volume fractions between 0.1% and 1.0% in the temperature range of 15–60 ^∘C. In the data set, temperature, volume fraction, and type of nanoparticles are considered as inputs, while thermal conductivity, dynamic viscosity, specific heat capacity, and density are considered as outputs. The hyper-parameters of the network were determined using the Bayesian optimization approach. Additionally, the k-fold cross-validation technique has been employed to prevent overfitting and improve the performance of the network. The optimum ANN structure results were compared with empirical correlations proposed by several authors. The findings indicate that the prediction capability of ANN, having a mean square error of 1.45e-4 and a coefficient of determination of 0.997265, outperforms that of correlations, enabling the straightforward prediction of multiple key properties of the studied water-based nanofluids through a single network rather than sophisticated correlations.