Predicting the Surface Tension of Liquids: Comparison of Four Modeling Approaches and Application to Cosmetic Oils.

Abstract

The efficiency of four modeling approaches, namely, group contributions, corresponding-states principle, σ-moment-based neural networks, and graph machines, are compared for the estimation of the surface tension (ST) of 269 pure liquid compounds at 25 °C from their molecular structure. This study focuses on liquids containing only carbon, oxygen, hydrogen, or silicon atoms since our purpose is to predict the surface tension of cosmetic oils. Neural network estimations are performed from σ-moment descriptors as defined in the COSMO-RS model, while methods based on group contributions, corresponding-states principle, and graph machines use 2D molecular information (SMILES codes). The graph machine approach provides the best results, estimating the surface tensions of 23 cosmetic oils, such as hemisqualane, isopropyl myristate, or decamethylcyclopentasiloxane (D5), with accuracy better than 1 mN·m-1. A demonstration of the graph machine model using the recent Docker technology is available for download in the Supporting Information.