Many compelling, selective catalysts are defined by metal-organic bonding around open shell transition metal centers. Although promising targets, the variable metal, ligand, oxidation, and spin states of these complexes present unique combinatorial and electronic structure challenges. I will describe the advances needed beyond the typical accuracy, efficiency, and approach of conventional density functional theory (DFT) workflows. I will discuss how, despite useful for addressing efficiency challenges, machine-learning(ML)-accelerated discovery both inherits the biases of DFT-derived training data and leads to attempted calculations of complexes that are not stable or feasible for DFT calculations. I will present an example from methane-to-methanol conversion where many calculations fail or have electronic structure properties (i.e., strong multireference or MR character) unlikely to be well predicted by DFT. I will describe how we have attempted to overcome these limitations by training ML models to predict multiple DFT functional results, enabling quantitative sensitivity analysis. I will describe how we have developed automated tools and trained ML models to predict the likelihood of calculation success with DFT and detect the presence of MR character. I will describe how these tools can be unified as the first steps toward autonomous workflows that avoid requiring expert intervention to determine when DFT-based catalyst discovery is robust.

Putting density functional theory to the test in machine-learning-accelerated discovery

ACS Spring Meeting 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.