Accelerating the Design and Discovery of Heterogeneous Catalysts Using Machine Learning

Abstract

In recent years, combined machine learning (ML) and quantum mechanical (QM) modeling approaches have emerged as a promising paradigm for accelerating the design and study of heterogeneous catalysts. Chapter 1 summarizes several recent applications and advances in the sub-field of ML in catalysis. Despite these advances, the field is nascent, with significant challenges to overcome to reach its full potential. The work presented in this dissertation directly addresses two of the challenges facing ML applications in heterogeneous catalysis: reducing the dependence of applications of ML in catalysis on black-box models and generating data-efficient catalysis datasets for high-throughput screening studies. Chapter 2–Chapter 4 focus on using intelligible ML techniques combined with QM modeling to uncover automated and intuition-agnostic descriptors and scientific insights. We introduce two works that use interpretable ML approaches to extract insights regarding chemisorption on subsurface alloy surfaces. The first, presented in Chapter 3, introduces a theory-guided ML approach that leverages tree-based generalized additive models and QM calculations to discover predictive geometric-structure-property chemisorption models. Interpretation of the models revealed three critical geometric features of the adsorption site that impact the relative chemisorption strength on metal alloys: the strain in the surface layer, the number of d-electrons in the ligand metal, and the size of the ligand atom. The second, presented in Chapter 4, uses an unsupervised machine learning approach leveraging principal component analysis to uncover geometric-structure-electronic-structure-property chemisorption models. The principal component descriptors connect well to prior efforts to develop electronic-structure descriptors and are primarily descriptive of the degree of orbital overlap at the surface and the number of valence d-electrons in the surface metal. We show that these effects can be well-captured by two atomic characteristics of the alloys' constituent metals, namely the surface and subsurface atoms' number of d-electrons and sizes. Notably, the work presented in both Chapter 3 and Chapter 4 yielded convergent insights regarding the roles of atomic size and d-electron character in governing the chemisorption properties of subsurface alloys. Chapter 5 focuses on generating data-efficient catalysis datasets using Bayesian optimization, with application to the screening of Ir-based oxides for the oxygen evolution reaction in acidic media. The screening identified Mo as a promising dopant for forming acid-tolerant Ir-based oxides, which were synthesized experimentally and confirmed to exhibit superior activity and stability to an Ir oxide control. This work highlights the promise of ML to guide the experimental exploration and optimization of catalytic materials in a data-efficient manner. Many open challenges remain in designing and understanding heterogeneous catalysis, and the approaches and tools developed in this dissertation are expected to streamline work in both due to their general applicability to any heterogeneous catalytic systems of interest, including but not limited to carbides, sulfides, nitrides, and intermetallic alloys. The dissertation concludes by outlining the natural research extensions of the work presented herein in Chapter 6.