A practical hybrid machine learning method for predicting the flash point of complex ternary alcohol-based mixtures

Abstract

The flash point (FP) modeling of ternary mixtures containing partially miscible regions, even by known models, is still challenging. This work proposes a practical and easy-to-use hybrid machine learning method, which combines artificial neural networks (ANNs) with the genetic algorithm (GA), to calculate the FPs of complex alcohol-based mixtures, particularly aqueous solutions of low-carbon alcohols. For this purpose, multilayer perceptron (MLP), radial basis function (RBF), and group method of data handling (GMDH) ANNs, were trained using the FPs of totally miscible region while the FPs of the partially miscible region were predicted by the trained ANN. By applying this approach, the FPs of partially miscible region of water-ethanol-1-butanol, water-ethanol-2-butanol, water-1-butanol-2-butanol, and methanol + toluene + 2,2,4-trimethylpentane were predicted. Another method was implemented by which the FPs of water-1-butanol-2-butanol mixture were predicted by only those of other mixtures, i.e., water-ethanol-1-butanol and water-ethanol-2-butanol. To the best of our knowledge, two applied approaches have not been previously used. Moreover, the FPs of two totally miscible aqueous solutions, namely water-methanol-ethanol and water-ethanol-isopropanol, were evaluated by different proposed ANNs in a usual way. The highest and lowest absolute average errors obtained by the applied methods for testing data were 5.09 and 0.45 °C, respectively.