Predicting the Self-Diffusion Coefficient of Liquids Based on Backpropagation Artificial Neural Network: A Quantitative Structure–Property Relationship Study

Abstract

The self-diffusion coefficient of pure liquids, a fundamental transport property, is involved in a wide range of applications. Many methods have been employed to study the self-diffusion coefficient, with the most popular being semiempirical models. The quantitative structure–property relationship (QSPR) has been widely used to predict various physicochemical properties of substances, but the appropriate molecular descriptors must be selected first. In this study, the charge density distribution area of molecules at a specific interval (Sσi) and cavity volume (VCOSMO) was determined based on the conductor-like screening model for the segment activity coefficient (COSMO-SAC). Using these molecular descriptors, a backpropagation artificial neural network (BP-ANN) method was employed to construct a nonlinear QSPR model that can predict the self-diffusion coefficients of pure liquids under normal pressure. The data set used included 2596 data points for 238 compounds, covering a self-diffusion coefficient range of 8.74 × 10–13 to 8.66 × 10–9 m2·s–1 and a temperature range of 90.5–475.1 K. The coefficients of determination (R2) of the BP-ANN model on the training, validation, and testing sets were all greater than 0.99. For the entire data set, the R2, absolute average relative deviation (AARD), and root mean square error (RMSE) were 0.9940, 7.09%, and 0.1106, respectively. In an application domain (AD) analysis, 94.67% of the data were within the AD range of the model. Consequently, the model developed in this study can satisfactorily predict the self-diffusion coefficients of liquids.