Advances in computing power and algorithms are increasingly enabling the high-throughput screening of novel materials using first-principles methods such as density functional theory (DFT). However, open shell molecular transition metal (TM) complexes present unique challenges to simulations owing to complicated spin- and oxidation state dependent behavior, sensitivity to functional used, a lack of reliable empirical methods and frequent calculation failure or convergence to unintended geometries and electronic states. These problems are exacerbated by the inherently combinatorial and effectively infinite size of potential TM complex space and the lack of large databases and the limited understanding of chemical similarity in TM space. We have developed a machine learning (ML) framework for TM complexes that begins to address all of these ideas: based on information from the molecular graph (i.e. connectivity) only, we have demonstrated prediction of spin state ordering to baseline DFT uncertainty levels as well predictions of orbital energies, redox potentials and catalytic energies. We have also proposed novel metrics of model uncertainty that allow us to judge the chemistry-dependent confidence in our ML models, and combined these with evolutionary algorithms to search design spaces for new complexes with targeted spin state ordering and frontier orbital properties that are validated at a DFT level the majority of the time. This approach represents at least an order-of-magnitude acceleration over naïve screening with DFT. However, ML can address other unique challenges: by training on different fractions of exact exchange, we can conduct design in a functional-dependent manner and predict the sensitivity of spin state ordering to functional choice. We also make predictions of optimal TM complex geometry, substantially improving on force-field based metal-ligand bond initialization. Analysis of the important parts of the molecular graph using statistical techniques can help understand structure-property dependence and similarity for TM complexes. Finally, we predict the likelihood of success of calculations to avoid performing unproductive simulations, greatly increasing the success rate of screening and avoiding waste of computational resources. All of these advances are distributed in open-source python packages molSimplify and molSimplify Automatic Design.

Beyond property prediction: What can machine learning do for virtual screening of transition metal complex space?

2019 Fall National Meeting

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