Mixture models inspired by the Kolmogorov-Arnold representation theorem

Abstract

Physical property models were developed to predict temperature-dependent multicomponent data using only temperature-independent binary parameters and pure component property temperature dependence. The Kolmogorov-Arnold representation theory was used to extend the linear blending rules and the Padé-like expressions describing the variation of physical properties of ideal solutions with composition. The effectiveness of correlating density, viscosity, refractive index and surface tension using this concept was tested. Ten ternary systems at either three or four different temperatures were regressed and compared to an ideal solution case. It was found that the four-parameter Kolmogorov-Arnold (KA) model performed excellently when the data regression included the full datasets. Unfortunately, the KA model may be too flexible, leading to overfitting binary data when applied to predicting ternary data.