Computational approaches and tools for modeling biomass pyrolysis

Thesis (Ph.D.)--University of Washington, 2016-08

Pyrolysis of lignocellulosic biomass is a promising process capable of producing renewable fuels and chemicals that are currently derived from nonrenewable sources. However, industrial pyrolysis processes to make these products from biomass are not yet economically viable and require significant optimization before they can contribute to existing oil-based transportation and chemical systems. One means of optimization uses kinetic and transport models for predicting the products of biomass pyrolysis, which serve as the basis for designing pyrolysis reactors capable of producing the highest value products. The goal of this work is to improve upon current pyrolysis models using computational fluid dynamics, detailed kinetic schemes, and machine learning. First, we develop a comprehensive two-dimensional particle model for wood pyrolysis that incorporates a multi-step semi-global reaction mechanism, prescribed particle shrinkage, and thermophysical properties that vary with temperature, composition, and orientation. This model is used to validate a new experimental technique that is capable of measuring the carbohydrate composition during fast pyrolysis under industrial conditions for the first time. Motivated by the challenges and limitations we encountered while developing this particle model, we next present a detailed kinetic model capable of predicting the temporal evolution of molecules and functional groups during lignin pyrolysis. This kinetic model provides information beyond the lumped yields of common pyrolysis models without any fitting, allowing it to cover a wider range of feedstocks and reaction conditions than lumped models. We perform an exhaustive sensitivity analysis using over two million simulations to explore which reactions contribute most to variance in the model predictions. Due to their computational expense, detailed kinetic models cannot be incorporated into comprehensive particle models that account for both kinetics and transport effects. To address this problem we demonstrate a machine learning approach to kinetic model reduction using neural networks that reduces the computational cost of our detailed kinetic model by four orders of magnitude. The trained neural networks generalize very well, predicting the outputs of the kinetic model with over 99.9% accuracy on new data.