(662a) Machine Learning for Homogeneous Open-Shell Transition Metal Catalyst Discovery

I will describe our efforts to accelerate the discovery of single site catalysts and materials. Despite decades of effort, no earth-abundant homogeneous catalysts have been discovered that can selectively catalyze the partial oxidation of methane to methanol. We have used our open source toolkit molSimplify to accelerate the discovery of candidate catalysts with machine learning. We exploit active learning to simultaneously optimize methane activation and methanol release calculated with machine learning (ML)-accelerated density functional theory (DFT) in a space of 16M candidate catalysts including novel macrocycles. In open shell transition metal catalysis, we show how conventionally used scaling relations are easily disrupted, motivating the development of independent ML models to predict multiple reaction energies that can be independently optimized. I will demonstrate how this approach accelerates the discovery of new design principles by at least a few orders of magnitude, leading to conclusions that are often unexpected. By constructing macrocycles from fragments inspired by synthesized compounds, we ensure synthetic realism in our computational searches. This large-scale search reveals that low-spin Fe(II) compounds paired with strong-field (e.g. P or S-coordinating) ligands and negatively charged axial ligands have among the best energetic tradeoffs between hydrogen atom transfer (HAT) and methanol release needed for methane to methanol conversion. This observation contrasts with prior efforts that have focused on high-spin Fe(II) with weak-field ligands. Time permitting, I will also discuss our analysis of complexes that have been previously synthesized and their relation to our design principles from the active learning approach.