Discovering transition metal complexes and metal organic framework catalysts with machine learning

Heather J. Kulik

Department of Chemical Engineering, Department of Chemistry, Massachusetts Institute of Technology, Cambridge, MA 02139

Machine learning accelerated density functional theory (DFT) has provided valuable insight into transition metal complex and catalyst properties. I will describe how we have used this approach to discover transition metal complexes as catalysts from millions of candidates for difficult chemical reactions such as the direct conversion of methane to methanol. Nevertheless, constructing a novel transition metal complex requires knowledge of how ligands will bind to metal centers and in what way those ligands will influence the properties at the metal center. I will describe how we have built models to predict ligand binding, including the likelihood of a ligand to favor hemilability (i.e., binding in multiple denticities). I will show how analysis of these models reveals unexpected trends going beyond heuristics for what influences ligand binding behavior. Since the orientation of ligands around a transition metal center can be varied significantly to lead to novel complexes, a small number of ligands (e.g., hundreds) can lead to a much higher number of theoretical complexes (e.g., trillions). I will describe both analytical and machine learning approaches we have developed to accelerate the discovery of lower symmetry complexes from data obtained for only the highest symmetry complexes. In the second half of my talk, I will describe how we have leveraged literature data to train machine learning models capable of predicting stability in metal-organic frameworks. I will describe how we have paired these models with computational screening with DFT and molecular modeling to discover robust materials with optimal properties for catalysis and separations.