The induction period (IP, obtained via standard conditions and a commercial instrument) is nowadays one of the most common tools to investigate oxidative stability in lipid samples. Despite the advantages and widespread use of this methodology in industry, the development of algorithms to predict IP values has been (thus far) impaired by the scarcity of data linking the composition of the sample with the corresponding IP value. Addressing this need and considering the persistent difficulties in bridging experimental and theoretical approaches related to edible fats, this paper reports on the development of a machine learning approach based on extreme gradient boosting (XGBoost) that was trained using a biodiesel database and then fine-tuned using a vegetable oils database, in both cases linking composition their corresponding stability (via rancimat). The resulting model was able to predict accelerated oxidative stability of vegetable oils with a mean average error of only 1.37 h, one of the lowest ever reported. Beyond demonstrating the ability of our model to provide meaningful predictions, this report highlights the importance of incorporating oxidation-related chemical descriptors and the utility of using associated databases and transfer learning to support the development of accurate machine learning models. • Machine learning approach to predict rancimat stability of vegetable oils • Transfer learning applied to fine tune a primary model trained on biodiesel data • Fatty acid composition used to make predictions of olive oil oxidative stability with a mean average error of only 1.37 h • Results underscore the importance of transfer learning to overcome data scarcity
Dike et al. (Wed,) studied this question.