A Hybrid Machine Learning Framework for Interpretable Kinetics of α-Tocopherol and Myricetin Synergism

Predicting the stabilizing efficacy of antioxidant mixtures in food oil emulsions is highly complex due to synergistic or antagonistic interactions between individual antioxidants. To address this challenge, we present an innovative hybrid machine learning framework, known as universal differential equations (UDEs), which integrates the expressive power of deep learning with the mechanistic boundaries of traditional kinetic models. We demonstrate the utility of this data-efficient approach by characterizing the coupled degradation dynamics of α-tocopherol in the presence of myricetin in oil. By embedding compact artificial neural networks directly into a system of ordinary differential equations, the hybrid UDE model successfully learned the hidden interactions from a small dataset, quantitatively revealing their mutualistic protection. Furthermore, we translated these machine-learned interactions into an interpretable, fully analytical model based on Hill-type saturation kinetics. Critically, this transparent analytical model not only accurately reproduced the training data (R² = 0.998) but successfully extrapolated antioxidant dynamics to previously unseen experimental formulations (R² = 0.978) with a fivefold increase in myricetin concentration. This work provides a powerful, interpretable AI tool for understanding complex kinetic interactions in food systems, with broad applications for accelerating product development, optimizing preservation strategies, and extending food shelf life.