Prediction of heterogeneous catalyst properties using data-driven multiscale modeling and software development

The world’s demand for energy and materials is increasing as the population and the average standard of living grows. Scaling up existing chemical processes is not a sustainable method to meet this demand as this will accelerate environmental issues such as climate change. Instead, novel catalysts that show better performance than the current state-of-the-art catalysts will be necessary to improve chemical process efficiency and reduce environmental impact. This progress, however, can be stifled if the fundamental understanding of catalysts is not prioritized. ☐ In this thesis, methods and software are developed to improve existing frameworks in multiscale modeling and are applied to heterogeneous catalysis systems with impact in environmental and energy applications. In Chapter 2, we show a methodology to improve biomass hydrodeoxygenation (HDO) activity over moderately reducible metal oxides. Previously, it was shown that the bulk formation energy of the metal oxide influences the activity but usually as a detriment to stability. We overcome this limitation by doping a relatively stable but inactive metal oxide (TiO2 anatase) with ultra-low loading catalyst (Pt). Through a combination of microkinetic modeling (MKM) and experimental kinetic studies, we find a significant increase in HDO activity compared to pure anatase without compromising the bulk stability or promoting unselective ring hydrogenation with only 0.04 wt.% loading. Further density functional theory (DFT) calculations, analysis of the MKM and catalyst characterization (STEM, EPR, FITR, XAS) reveal the dopant promotes the formation of oxygen vacancies. ☐ In Chapter 3, we develop a new model called the lattice convolutional neural network (LCNN) model to calculate coverage effects between adsorbates. As a case study, we demonstrate its performance on a O/NO/Pt(111) system, which is important for the operation of the three-way catalyst present in vehicle exhausts systems. We find the LCNN performs better than the state-of-the-art models including a cluster expansion trained with the genetic algorithm and the convolution operation of the crystal graph convolutional neural network by 20 – 30% (test RMSE of 4.4, 5.5, and 6.8 meV/site, respectively). Furthermore, we analyze the van der Waals interactions by visualizing hidden representation of the adsorbate lattice system in terms of individual site formation energies. ☐ In the last two chapters, we show software developed to streamline conversion of microscopic properties to mesoscopic properties and generation of volcano curves. We present the Python Multiscale Thermochemistry Toolbox (pMuTT), a Python library specialized for thermodynamic and kinetic parameter estimation. It houses statistical mechanical models to convert ab-initio data to thermochemical properties and supports many empirical and semi-empirical relationships used in the field including NASA polynomials, Bronsted-Evans-Polanyi (BEPs) relationships, and linear scaling relationships (LSRs). We also present the Descriptor-based Microkinetic Analysis Package (DescMAP), which is Python software for automated descriptor selection and volcano curve generation. It leverages existing software in the Virtual Kinetic Laboratory (VLab) suite for enhanced functionality and uses easily modifiable spreadsheets and template scripts for great customizability. We demonstrate its capabilities by generating volcano curves for the nonoxidative dehydrogenation of ethane to ethylene, an important process in the chemical industry. We compare volcano curves generated using atomic descriptors versus descriptors selected by the software and find the software’s descriptors produce more well-defined volcano curves.