(192b) First-Principles Design of Heterogeneous Catalysts with Automatic Simulations

Producing high-valued commodities and energy-storing molecules from non-petroleum sources is among the challenges of decarbonizing the economy by 2050. However, known catalysts are often expensive, display poor reaction kinetics, or are limited in synthetic accessibility. Discovering cost-effective materials for clean energy reactions is decisive to overcome pressing climate goals. In this talk, I will discuss how a wide range of computational methods can accelerate catalyst discovery. In particular, I will show how combining graph theory, high-throughput simulations, density functional theory (DFT), and machine learning (ML) can elucidate different aspects of materials discovery in thermal and electrocatalysis.

I first illustrate with the case of zeolite catalysts. Due to their topological diversity, zeolites suffer from phase competition and transformations, which directly impact their stability. Using graphs to model atomic rearrangements, I demonstrate which topologies are related by diffusionless transformations and intergrowths, reproducing results from the literature and proposing new synthesis pathways for zeolites. Then, to quantify phase competition in zeolite synthesis, I developed a computational platform to predict host-guest interactions two orders of magnitude faster than the existing methods. By simulating millions of pairings between frameworks and organic structure-directing agents (OSDAs), binding metrics that explain phase selectivity in zeolites are proposed. The theory reproduces outcomes from more than one thousand papers extracted from the literature, and enables the design of selective, chemically simple OSDAs through a web-based design platform. Experiments show that the computer-designed OSDAs significantly improve catalytic properties of zeolite frameworks and produce previously inaccessible structures. The kinetic mechanisms of confined reactions are further studied using DFT, elucidating the origins of such enhancement in catalytic properties and introducing new design principles for OSDA selection. Finally, machine learning approaches are used to explore the space of OSDAs towards millions of possible molecules. These works can enable yet unrealized zeolites to be accessed, and reactions previously unfavorable to be catalyzed by novel nanoporous materials.

A second example is shown with multicomponent transition metal oxides. In contrast with zeolites, these materials are limited in structural diversity, but display a combinatorial number of realizable compositions. To simultaneously optimize their activity and stability, automated DFT simulations are used to map composition to electronic structure descriptors, which are then interpreted using atomic features. Literature data shows that DFT-based stability correlates with synthesis outcomes of these materials, thus validating the design strategy. The data is then used to train ML models, providing faster evaluation of electronic properties and reduction of prediction uncertainties in an active learning loop. Through an interplay between experiments and theory, materials selection strategies are successively refined, leading to experimental realization of potentially superior structures in terms of activity and stability. This iterative approach provides a roadmap to improving electrocatalysts from theory-driven research.