Machine Learning-Assisted Design of Molecular Structure of Diphenylamine Antioxidants

This study calculated the bond dissociation energy (BDE), solubility parameter (δ), binding energy (E_binding), and other parameters of 96 diphenylamine antioxidants using molecular simulations to verify the quantitative relationship between the structure and properties of antioxidants and obtained 288 antioxidant performance parameters to build a machine learning (ML) data set. A group-partitioning scheme consisting of 10 group descriptors and 3 molecular connectivity chi index descriptors was established. The antioxidant parameters were combined with the partitioning results to form a complete ML data set. An artificial neural network model (ANN) was used to quantize the structure–performance relationship of antioxidants. The correlation coefficient (R) between the predicted value and the true value was greater than 0.84, the mean relative error (ARE) of BDE was less than 4.53%, the ARE of δ was less than 1.21%, and the ARE of E_binding was less than 14.76%. The random forest (RF) model was used to study the contribution of each descriptor to oxidation resistance and analyze the effects of different substituent locations. The results of the chemical and physical analyses showed that the introduction of alkyl chains improved the performance of the antioxidants. An alkyl chain with nine carbons in the main chain was grafted to site 4 as a new substituent, and a new molecular structure of diphenylamine antioxidants was designed. Compared with the skeleton structure, the BDE of the newly designed antioxidant decreased by 1.18%, δ decreased by 6.11%, and E_binding increased by 83.44%. This demonstrates the effectiveness of the ML-assisted molecular design of antioxidants.