Carnegie Mellon University
Broderick_cmu_0041E_11131.pdf (52.81 MB)

Untitled ItemBridging Computation and Experiment in Heterogeneous Catalysis with Applied Machine Learning

Download (52.81 MB)
posted on 2024-04-12, 19:02 authored by Kirby Broderick

 With the growth of computational resources and methods of the past fifty years, simulations have become a useful scientific and engineering tool across many fields. From photonics to protein folding, computers are now routinely used to discover novel materials. The decline in energy costs due to renewable energy breakthroughs is paving the way for technologies that exploit cheap, clean energy, hydrogen production foremost among them. Advances in heterogeneous catalysis are necessary to commercialize this technology, and further developments in this field promise to decarbonize thorny industries such as transportation and chemical processing and also enable efficient CO2 capture. 

However, catalyst discovery still largely proceeds through trial and error, with expert-guided experimentation supplemented by computational mechanistic investigations. Density Functional Theory has been commonly used to model catalysts, but its cost and complexity has led researchers to develop machine learned surrogates that are currently approaching accuracies necessary for materials discovery. The main challenge that lies ahead is in figuring out how to align experiments and industrial processes with effective computational representations. 

In this thesis, I apply pretrained and finetuned Density Functional Theory-based surrogate models within a variety of computational workflows to address challenges in materials science. First, I apply a popular model of catalytic activity to a highly complex real-world experimental dataset of hydrogen-evolving catalysts. Although the model describes the most active experiments effectively, its inability to capture the behavior of the entire space led me to develop new tools to address surface stability. I developed a computational methodology to calculate the cleavage energy of crystal facets, demonstrating significant improvements in accuracy over the previous state of the art. This approach also showed much better discrimination between different terminations in the same crystallographic orientation, an important classification problem in materials design. Finally, I designed efficient and scalable software to conduct Monte Carlo surface segregation simulations in a high performance computing environment and demonstrated how finetuned large graph models can be utilized to predict the outcomes of surface segregation experiments in fcc ternary metal alloys. 




Degree Type

  • Dissertation


  • Chemical Engineering

Degree Name

  • Doctor of Philosophy (PhD)


John Kitchin