Interpretable Machine Learning with Physics-Based Feature Engineering: Application to the Catalytic Cracking Reaction of Polycyclic Aromatic Hydrocarbons

Machine learning has been widely
used to predict heterogeneous
catalytic reactions, but the trade-off between prediction accuracy
and interpretability remains an open issue. This study attempts to
develop an interpretable machine learning model for predicting the
products of heterogeneous catalytic reactions with complex feedstock
compositions. The target reaction was the catalytic cracking reaction
of polycyclic aromatic hydrocarbons, and the compositions of the reaction
products (monocyclic aromatic hydrocarbons, condensed two-ring aromatic
hydrocarbons, and heavy components) were predicted. The combination
of the least absolute shrinkage and selection operator (LASSO) and
physics-based feature engineering significantly improved the prediction
accuracy compared with that of the regression model with only basic
descriptors (reaction temperature, contact time, and feedstock composition).
The prediction accuracy of the constructed model exceeded that of
the black-box nonlinear regression models. Furthermore, it was confirmed
that information about important reactions could be extracted from
the magnitude of the standard regression coefficients. These findings
imply that a combination of LASSO and physics-based feature engineering
can be used to construct machine learning models with high prediction
accuracy and interpretability for heterogeneous catalytic reactions.