ChE PhD Dissertation Defense: Chao Xu

Name:
Chao Xu

Title:
Towards Automated Heterogeneous Catalysis Design: Integrating Activation Energy Estimation, Uncertainty Quantification, and Coverage-dependent Thermodynamics in Microkinetic Modeling

Date:
1/15/2025

Time:
12:00:00 PM

Committee Members:
Prof. Richard West (Advisor)
Prof. Qing Zhao
Dr. Franklin Goldsmith
Dr. Zack Ulissi

Location:
EXP 610-A

Abstract:
In the context of climate change, reducing CO2 emission and advancing sustainable development are global priorities. Developing efficient “green” fuel processes requires overcoming the low productivity of industrial catalysts, necessitating new catalyst designs. Traditional trial-and-error approaches are costly and time-intensive, but multi-scale modeling provides a low-cost, efficient alternative by exploring catalyst design across atomic to reactor scales.

This dissertation enhances scientific software such as Reaction Mechanism Generator (RMG) and Cantera etc. for automating heterogeneous catalysis modeling to accelerate catalyst design. An active-learning workflow for calculating coverage-dependent thermodynamics with EquiformerV2, a graph neural network, was developed, complementing the existing RMG coverage-dependent kinetics functionality. The estimated coverage-dependent thermodynamics were validated through a CO/H2 methanation model on Ni. Because reaction activation energies are also affected by species’ coverage, the dissertation also addresses rapid reaction barrier estimation by implementing the Blowers-Masel Approximation (BMA), which relates a reaction barrier to enthalpy using minimal data. Integrated into Cantera, this method supports catalyst screening with linear scaling relationships (LSRs) and BMA kinetics. A methane partial oxidation study on 81 hypothetical metals demonstrated how BMA affects rate-limiting species and high-selectivity catalysts, while sensitive reactions remain unaffected.

Lastly, thermodynamic properties of surface species in catalytic methane partial oxidation models on Rh were estimated using DFT, LSRs, and a graph neural network named GEMNET, combined with Bayesian parameter estimation for process optimization. The three methods provided close thermodynamic data with different prior uncertainties, validating the feasibility of combining machine learning potentials with RMG for thermodynamic estimation on all kinds of binding facets. Bayesian parameter estimation improved simulation accuracy of the three estimation methods while uncovering active site information inaccessible through conventional means. The workflow effectively integrates experimental and computational uncertainties, enabling data-informed catalyst design and optimization.