(191i) Accelerating the Design of Single-Site Materials for Catalysis Using Computational Data, Experimental Data, and Machine Learning

A revolution in data-driven machine learning (ML) methods has unlocked the design of new molecules and materials. Computational high-throughput virtual screening with first-principles density functional theory (DFT) plays a valuable role in unearthing design rules for scalable and viable synthetic single-site transition metal complex catalysts that preserve selectivity and activity observed only in metalloenzymes. These catalysts represent the most promising synthetic analogues to metalloenzymes, often enabling atom-economy and tunability not possible with bulk heterogeneous catalysts. The 3d transition-metals that comprise these catalysts access a range of oxidation and spin states, complicating analyses of their reaction energetics.

I will present a systematic analysis of the oxidation and spin state dependent behavior of single-site light alkane oxidation catalysts. In particular, I will show that scaling relations that have been proposed for screening partial methane oxidation catalysts do not hold generally when considering multiple metals, oxidation states, and spin states. These exceptions are some of the most interesting candidate materials to study for reactivity. I will illustrate how we can harness broken scaling relations to design catalysts with optimal property tradeoffs. First, I will highlight our efforts to design catalysts that are both oxidatively stable but favorably form metal-oxo moieties that are utilized for C-H activation. Then, I will show how assumptions of a strong thermodynamic linear free energy relationship for C-H activation leads to incorrect conclusions on intermediate and high-spin iron-oxo moieties that are typically invoked for catalysis. Lastly, I will show how we can harness oxidation and spin state dependent scaling to design catalysts with Pareto optimal tradeoffs between two distinct reaction energy steps, over a space of over 16 million candidate catalyst materials.

Metal-organic frameworks (MOFs) are solid-state materials that can have single-metal active sites in analogy to transition metal complexes. For nearly two decades, MOFs have been developed for applications including heterogeneous catalysis. In practice, we must activate a MOF and remove solvent from its pores to render it porous and usable. Simultaneously, the MOF must also be stable under the thermal conditions to be useful for catalysis. A lack of understanding of how to improve MOF stability limits their practical use. I will highlight our developments towards gaining insights for MOF stability. First, I will highlight our divide-and-conquer graph-theoretic MOF representations, which have enabled us to compare the diversity of hypothetical and experimental MOF databases. I will then show how we can use the insights from our MOF chemical space maps to construct a path forward for ML-driven MOF design. Since computation cannot readily predict thermal or solvent removal stability, we use natural language processing (NLP) to extract stability information over thousands of MOF manuscripts, to obtain insights on solvent removal and thermal stability of these materials. We use these models to construct a new hypothetical MOF database that is closer to experimental realism, has improved diversity, and contains MOFs that are likely to be stable.

In conclusion, my work demonstrates how combining DFT, experimental data, and ML can lead to the design of new single-site materials that are useful for catalysis applications. I will highlight how all aspects of my work come together to build an end-to-end materials platform that both reproduces existing chemical knowledge, while enabling and accelerating searches through multi-million molecule chemical spaces.