(365e) Evaluation of the Economic Viability and Sustainability of Autothermal Pyrolysis Sugars through Computation and Machine Learning

Autothermal pyrolysis (ATP) is a novel approach to converting biomass into chemicals, fuels, and bioproducts. ATP employs a small amount of oxygen to partially combust biomass and some pyrolysis products inside the reactor. This avoids heat losses associated with indirect heating and enables process intensification of pyrolysis reactors. Recent studies demonstrated the production of pyrolytic sugars, bioasphalt, and biochar from red oak and corn stover ATP. Experimental results indicate that ATP product yields are comparable to conventional pyrolysis while reducing the need for heat generation equipment. Feedstock properties could have a significant impact on the yield and quality of ATP products. This study employs computation and machine learning to investigate the impact of feedstock properties on ATP products and their economic and environmental sustainability.

Previous studies indicate that the costs and emissions of pyrolysis products can vary widely depending on the feedstock properties. For example, the minimum selling-fuel price (MFSP) of pyrolysis biofuels could range between $2 and over $5 per gallon. Several methods have been proposed to improve the prediction ability of pyrolysis computational models and reduce commercialization risk. However, few studies have employed machine learning models and techniques. The objective of this paper is to investigate the ability of machine learning models to predict the minimum sugar-selling price (MSSP) and greenhouse gas (GHG) emissions of producing sugars, biochar, and bioasphalt ATP.

This paper integrates machine learning alongside a chemical process model, techno-economic (TEA) and life cycle analysis (LCA) of the ATP system. Commercial biochemical feedstock composition was provided by Idaho National Laboratory (INL). Approximately 3000 cellulose, hemicellulose and lignin data samples were generated using a Generative Adversarial Network (GAN) machine learning model. A Kriging-based machine learning model predicted ATP product yields based on the GAN samples. Product yields were introduced into the chemical process model to generate mass and energy balance for TEA and LCA calculations. Preliminary results estimate an average MSSP of around $400/ tonne of sugars and net GHG emissions of -0.16 kg CO2eqv/MJ of sugar. These results demonstrate that machine learning models could help reduce the commercialization risk of biorefinery techonologies.