Data-driven mechanism derivation in hydrogenation reaction systems: A hybrid approach integrating machine learning and mathematical models

초록

This study proposes a hybrid approach that integrates a black-box model with an equation-based mechanism verification model to identify the causative reactants of side reactions occurring in hydrogenation reaction systems and to derive their mechanisms. Due to the complex reaction network inherent in hydrogenation systems, exhaustive exploration and prediction of all side reactions are practically difficult, and such predictive uncertainty can lead to deterioration in product quality and economic losses. Accordingly, this study constructed a reliable dataset by performing preprocessing steps to remove missing values and outliers based on actual operating data, and designed an Artificial Neural Network (ANN) model. As a result of enhancing the generalization performance of the model through data transformation that reflects domain knowledge in the form of reaction rate expressions, the coefficient of determination (R2) improved to 0.97, and the root mean squared error (RMSE) decreased markedly to 0.04. In addition, the influence of each reactant was quantified and visualized using eXplainable AI (XAI) techniques, based on which the side reaction mechanisms were derived. To verify the derived mechanisms, reaction rate expressions were formulated using the Langmuir–Hinshelwood (L–H) and Power Law models, and a Monte Carlo-based mathematical model was developed. The validation results showed very high performance for all products, with R2 values greater than 0.98 and mean absolute percentage error (MAPE) lower than 2.31%. This study has academic and industrial significance in that it systematically elucidates actual side reaction mechanisms while presenting a highly reliable predictive model applicable to reaction control design. © 2026 Elsevier B.V.

키워드