Machine learning-assisted design of the molecular structure of <i>p</i>-phenylenediamine antioxidants

This study employed machine learning to predict the solubility parameter (<i>δ</i>) and bond dissociation energy (BDE) of antioxidant molecules, focusing on <i>p</i>-phenylenediamine derivatives with varying carbon chain lengths, side group positions, and functional groups (-CH₃, -OH, and -NH₂). The multilayer perceptron (MLP) model, enhanced by data augmentation and genetic algorithms, was developed to correlate the "molecular structure-descriptor-target parameter" relationship. The model achieved high prediction accuracy (coefficient of determination >0.86, relative percent difference >2.62). SHapley Additive exPlanations analysis revealed molecular polarity as the key factor influencing antioxidant performance. Molecules with -NH₂ side groups exhibited lower BDE values. A <i>p</i>-phenylenediamine derivative with 'CH₃[CH₂]₁₃CH(NH₂)-' connected to an aniline group showed optimal properties (Δ<i>δ</i> = 0.02 (J cm^-3)^0.5, BDE = 289.46 kJ mol^-1). Molecular simulations confirmed that the proposed antioxidant has excellent compatibility, anti-migration, and antioxidant activity in triglyceride oil. This study demonstrates the utility of MLP models for designing high-efficiency antioxidants for edible oils.