Investigating the Impacts of Feedstock Variability on a Carbon-Negative Autothermal Pyrolysis System Using Machine Learning

Abstract

Feedstock properties impact the economic feasibility and sustainability of biorefinery systems. Scientists have developed pyrolysis kinetics, process, and assessment models that estimate the costs and greenhouse gas (GHG) emissions of various biorefineries. Previous studies demonstrate that feedstock properties have a significant influence on product costs and lifecycle emissions. However, feedstock variability remains a challenge due to a large number of possible feedstock property combinations and limited public availability of feedstock composition data. Here, we demonstrate the use of machine learning (ML) models to generate large feedstock sample data from a smaller sample set for sustainability assessment of biorefinery systems. This study predicts the impact of feedstock properties on the profitability and sustainability of a lignocellulosic biomass autothermal pyrolysis (ATP) biorefinery producing sugar, phenolic oil, and biochar. Generative Adversarial Networks and Kernel Density Estimation machine learning models are used to generate 3,000 feedstock samples of diverse biochemical compositions. Techno-economic and lifecycle assessments estimated that the ATP minimum sugar selling price (MSSP) ranges between $66/metric ton (MT) and $280/MT, and the greenhouse gas (GHG) range from a net negative GHG emission(s) of −0.56 to −0.74 kg CO2e/kg lignocellulosic biomass processed. These results show the potential of ML to enhance sustainability analyses by replacing Monte Carlo-type approaches to generate large feedstock composition datasets that are representative of empirical data.