Development of machine learning model for the prediction of selectivity to light olefins from catalytic cracking of hydrocarbons

Abstract

Light olefins are the primary building block for the production of petrochemicals and polymers. Light olefins are largely produced from steam/catalytic cracking of naphtha or ethane/propane. Selectivity to light olefins is significantly dependent on the reaction conditions. In this article, several machine learning models are developed and tested to predict the selectivity of ethylene and propylene using seven input features. For this study, a total of eight ML models consisting of adaptive boost, extreme gradient boost, categorical boost, light gradient boost, decision tree with bagging, random forest, k-nearest neighbour, and artificial neural models are developed. The extreme gradient boost model gave the highest prediction accuracy for the ethylene selectivity, while the light gradient boost gave the highest R² for the propylene selectivity. The SHAP analysis showed the input parameter’s importance ranking for ethylene predictions as temperature > number of carbon atoms > Si/Al ratio > acidity > weight hourly space velocity > effect of diluent > number of hydrogen atoms. The importance ranking of input parameters for propylene selectivity was observed as weight hourly space velocity > acidity > temperature > Si/Al ratio > effect of diluent > number of carbon atoms > number of hydrogen atoms.