(201b) Predictive Machine Learning for Thermal Depolymerization of Plastics and Application to Economic Analysis of Municipal Recycling Facilities

Chemical recycling via thermal processes such as pyrolysis and hydrothermal liquefaction (HTL) is a potentially viable way to convert mixed streams of waste plastics into usable fuels. Unfortunately, experimentally determining product yields for real waste streams can be time and cost prohibitive, especially since the yields are very sensitive to both feed composition and reaction conditions. Data driven methods capable of predicting yields allow for rapid analysis of a specific feedstock’s potential for chemical recycling. In previous work¹ a machine learned model capable of predicting oil yields from pyrolysis reactions was developed and used to demonstrate the thermodynamic potential of such a process with real waste. However, this model had a few key limitations, including being applicable only to pyrolysis reactions, and not HTL; restriction to the big six polymers (high- and low-density polyethylene, polypropylene, polystyrene, polyethylene terephthalate and polyvinyl chloride); and not being able to handle reaction/residence time as a model feature. This work aims to address these limitations and develop a more widely applicable model capable of predicting reaction oil yields for both pyrolysis and HTL reactions. The data set used to develop the previous model was expanded to include HTL data without adding more features by converting the polymer classification from mixture composition of the big six polymers to elemental composition of the feed and molecular weight of the plastics monomers contained within it. The inclusion of the molecular weight of the monomer was necessary for differentiating polymers such as polyethylene and polypropylene that have the same elemental composition. This change in classification allows for the inclusion of polymers outside the big six without adding many new features which would risk overfitting while increasing the data set size, particularly for HTL reactions. The updated data set includes 446 data points (increased from 325 previously). This data set also includes reaction time and residence time as features, by allowing both to exist regardless of reactor type. The dataset was used to train and optimize a Random Forest model, which resulted in a test set mean absolute error (MAE) of ± 11.5%. The new model can be used for a wide range of plastics and is applicable to both pyrolysis and HTL depolymerization technologies. Here, the model was used to predict the oil yield possible from plastic wastes reported from Municipal Recycling Facilities (MRFs) from a variety of locations. These predictions were used to perform a techno-economic analysis (TEA) for conversion of plastic waste to pyrolysis oil. Data driven models such as these are important decision-making tools allowing for the assessment of new feedstocks and the allocation of finite resources.

1. E. R. Belden, M. Rando, O. Ferrara, E. Himebaugh, C. Skangos, N. K. Kazantzis, R.C. Paffenroth, M.T. Timko. Machine learning predictions of oil yields obtained by plastic pyrolysis and application to thermodynamic analysis. ACS Engineering Au, 2022