Machine Learning Predictions of Oil Yields Obtained by Plastic Pyrolysis and Application to Thermodynamic Analysis

Abstract

Chemical recycling via thermal processes such as pyrolysis is a potentially viable way to convert mixed streams of waste plastics into usable fuels and chemicals. Unfortunately, experimentally measuring product yields for real waste streams can be time- and cost-prohibitive, and the yields are very sensitive to feed composition, especially for certain types of plastics like poly(ethylene terephthalate) (PET) and polyvinyl chloride (PVC). Models capable of predicting yields and conversion from feed composition and reaction conditions have potential as tools to prioritize resources to the most promising plastic streams and to evaluate potential preseparation strategies to improve yields. In this study, a data set consisting of 325 data points for pyrolysis of plastic feeds was collected from the open literature. The data set was divided into training and test sub data sets; the training data were used to optimize the seven different machine learning regression methods, and the testing data were used to evaluate the accuracy of the resulting models. Of the seven types of models, eXtreme Gradient Boosting (XGBoost) predicted the oil yield of the test set with the highest accuracy, corresponding to a mean absolute error (MAE) value of 9.1%. The optimized XGBoost model was then used to predict the oil yields from real waste compositions found in Municipal Recycling Facilities (MRFs) and the Rhine River. The dependence of oil yields on composition was evaluated, and strategies for removing PET and PVC were assessed as examples of how to use the model. Thermodynamic analysis of a pyrolysis system capable of achieving oil yields predicted using the machine-learned model showed that pyrolysis of Rhine River plastics should be net exergy producing under most reasonable conditions.