Predictive large reaction network modeling via data-driven methods: application to biomass conversion
| Author(s) | Gu, Geun Ho | |
| Date Accessioned | 2018-09-20T11:38:33Z | |
| Date Available | 2018-09-20T11:38:33Z | |
| Publication Date | 2018 | |
| SWORD Update | 2018-07-24T13:06:53Z | |
| Abstract | US energy demand is projected to continue increasing, and exploiting sustainable resources is critical to minimizing risks in economy, geopolitics, and environment. Lignocellulosic biomass is a key sustainable source of carbon for the production of renewable fuels and chemicals. Compared to petroleum processes, biomass utilization requires selective de-functionalization, calling for innovations in catalysis. Previously, metal surface catalysts have shown promising performance for hydrogen generation and hydrodeoxygenation (HDO) in biomass conversion processes. However, mechanistic understandings are lacking and the search for optimal catalysts continues. In this regard, density functional theory (DFT) studies can aid, but DFT is computationally too expensive to investigate large reaction networks of biomass monomers on metal surfaces. Previous research introduced a semi-empirical based framework to reduce the computational cost and rapidly build microkinetic models (MKMs), but several tasks remain to realize the framework. This thesis aims to narrow the gaps in this framework. ☐ One gap is modeling capability for lignin monomer derivatives whose theoretical investigation has been minimal prior to this thesis. This thesis has two overarching goals. First, we build reliable semi-empirical methods mainly for thermochemistry and to an extent for kinetics. Second, we apply these methods to the HDO mechanism of a lignin model compound, cresol, on a metal surface for the first time. ☐ Three chapters are dedicated to improving and expanding the capability of thermochemistry prediction for molecules on metal surfaces. With the reaction network identified, group additivity for lignin monomers is developed. The new group additivity introduces more sophisticated descriptors based on the theoretical basis of the group additivity and electron density analysis. The cross-validation shows 2.8 kcal/mol mean absolute error (MeanAE) with 591 data points, an improvement over the previous framework by 2.3 kcal/mol in the MeanAE. The model is capable of rapidly predicting ~14,000 molecules (accounting different binding geometries) in the reaction network of a lignin model compound, guaiacol. Next, group additivity for solvated molecules on a metal surface is developed that provides an excellent first approximation of solvation energy at MeanAE of 1.0 kcal/mol at no computational cost. To further improve the thermochemistry prediction framework, a machine learning method called LASSO was introduced. LASSO is used to automatically select the most informative patterns (descriptors) of the molecules in datasets. The application to the 591 lignin dataset reveals a MeanAE of 2.08 kcal/mol, achieving the sub 0.1 eV error for the first time for an adsorbate group additivity method. In order to simplify the user interface for these thermochemistry prediction methods as well as to automatically build datasets, a machine-learning algorithm for predicting adsorption geometries is presented. Methods are made available on Github. ☐ Next application of these methods on MKM is carried out. Our DFT-refined MKM of cresol on Pt(111) demonstrates a novel mechanism consistent with multiple experiments, suggesting that Pt alone is capable deoxygenation. The MKM and DFT results suggest that cresol undergoes deoxygenation upon sufficient ring hydrogenation. Innovations in the semi-empirical model frameworks are made to build the MKM. In order to rapidly estimate activation barriers of ~500 reactions, Brønsted-Evans-Polanyi (BEP) relations were developed. The previous approach for building BEP relations results in a significant MeanAE of 17.3 kcal/mol for the rate-limiting C-OH scission reactions. A new framework introduced herein categorizes transition states based on their structure and reduces the MeanAE to 1.6 kcal/mol. In addition, the computational cost in building a model for lateral interactions has been reduced by coarse-graining surface species (from 150 to 7 species). | en_US |
| Advisor | Vlachos, Dionisios G. | |
| Degree | Ph.D. | |
| Department | University of Delaware, Department of Chemical and Biomolecular Engineering | |
| Unique Identifier | 1053623186 | |
| URL | http://udspace.udel.edu/handle/19716/23800 | |
| Language | en | |
| Publisher | University of Delaware | en_US |
| URI | https://search.proquest.com/docview/2088493016?accountid=10457 | |
| Keywords | Applied sciences | en_US |
| Keywords | Biomass | en_US |
| Keywords | Conversion | en_US |
| Keywords | Data-driven | en_US |
| Keywords | Kinetic | en_US |
| Keywords | Machine learning | en_US |
| Keywords | Predictive | en_US |
| Title | Predictive large reaction network modeling via data-driven methods: application to biomass conversion | en_US |
| Type | Thesis | en_US |