Machine-learning models for analysis of biomass reactions and prediction of reaction energies

Biomass and derived compounds have the potential to form the basis of a sustainable economy by providing a renewable source of many chemicals. The selective synthesis and conversion of biomass compounds are often catalyzed by transition metal catalysts. Computational screening has emerged as a promising tool for discovery and optimization of active and selective catalysts, but most existing examples focus on small molecule reactions. In this study, the density functional theory (DFT) approach is first validated by comparing computational results to experiments for ethanol conversion over molybdenum oxide. Subsequently, DFT is combined with machine-learning approaches to identify and overcome challenges associated with computational screening of biomass catalysts. A recursive algorithm is used to elucidate possible intermediates and chemical bond cleavage reactions are for linear biomass molecules containing up to six carbons. Machine-learning algorithms based on the Mol2Vec embedding are applied to classify reaction types and predict gas-phase reaction energies and adsorption energies on Rh(111) (MAE ~0.4 eV). With the workflow, we are able to combine the physics-based density functional tight binding method with the machine learning model to identify the stable binding geometries of biomass intermediates on the Rh (111) surface. Finally, we show preliminary results toward the development of a neural network force field based on the Gaussian multipole feature approach. The results indicate that this strategy is a promising route toward fast and accurate predictions of both energies and forces of hydrocarbons on a range of transition-metal surfaces. The results of this thesis demonstrate the utility of machine-learning techniques for studying biomass reactions, and indicate the potential for further developments in this field.

Date

2021-11-22

Resource Type

Text

Resource Subtype

Dissertation

Full item page