Combining Cheminformatics and Machine Learning to Accelerate Open-Shell Catalyst and Materials Discovery

The discovery of new catalysts and materials is critical for energy conversion, energy storage, and molecular separations, yet sifting through large chemical spaces of hundreds of thousands to millions of molecules to uncover the best candidates is often a major bottleneck. Virtual high-throughput 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. DFT alone, however, cannot be used to sift through large design spaces due to its computational cost. When DFT data is too expensive to be gathered in large quantities, we must consider alternative data sources such as curated experimental data.

In this talk, I will present my efforts to address challenges in efficiently screening single-site materials. In the first part of my talk, I will present a systematic analysis of the oxidation and spin state dependent behavior of open-shell single-site transition metal catalysts for partial methane oxidation to methanol. I will discuss how the limits of existing catalyst screening tools (e.g. linear free energy relations (LFERs)) can be overcome by using machine learning (ML) over multi-million molecule chemical spaces, allowing the discovery of best-in-class catalysts. In the second part, I will highlight my efforts to build a toolkit to screen metal-organic frameworks (MOFs) in analogy to transition metal complexes. I will demonstrate how combining new graph theoretic representations of MOFs with experimental data curated by natural language processing (NLP) leads to new insights on solvent removal and thermal stability of these materials when computation cannot be used. Overall, this work provides a paradigm for combining DFT, experimental data, and ML to accelerate the design of single-site inorganic catalysts and materials across different length scales.