Discovery of ester lubricants with low coefficient of friction on material surface via machine learning

Abstract

The design of lubricants with excellent anti-wear performance is the focus of current tribological research. The anti-wear performance of lubricants is fundamentally determined by their chemical structures. Thus, the development of efficient quantitative structure–property relationship (QSPR) to accurately predict lubricant’s anti-wear performance can be a powerful approach toward the design of new lubricants. To this end, this work aims to establish an efficient QSPR model using Least Squares-Support Vector Regression (LS-SVR) method, which employs the screened important molecular descriptors as the input to represent the characteristics of lubricants. Specifically, 1444 molecular descriptors of 54 ester molecules are computed based on the molecular graphics and adjacency matrix. Multiple stepwise regression (MSR) algorithm is used to select important molecular descriptors, and 5 important and unrelated molecular descriptors are used as input in LS-SVR. The mean relative error (MRE) of the training set, test set and external set from the established QSPR model are 2.92%, 6.03% and 9.30% compared with experimental values, respectively, indicating the developed QSPR model exhibits good predictive ability in CoF values of certain molecules and can be used to predict the anti-wear performance of lubricants only through the 5 important molecular descriptors. In addition, it is shown that the anti-wear performance of lubricants can be enhanced by increasing the longest aliphatic chain, the relative molecular mass and van der Waals volume of the studied lubricant molecules via MSR analysis. In the end, we screened the compounds containing ester groups in the database and obtained 8 potential lubricants, which provided guidance for the discovery of new lubricants.