Employing machine learning models for high throughput screening of single-site molecular catalyst energetics

Computation and simulation have emerged as essential tools for accelerating the discovery of catalysts and rationalizing their performance. Within heterogeneous catalysis, the linear free energy relationship (LFER) between energies of key intermediates or transition states in a catalytic cycle is a well-established tool employed to drastically reduce the computational cost of high throughput catalyst screening. The validity of LFERs in single-site catalysts, which are promising avenues to enhanced selectivity and activity, has been shown to be limited, owing to the greater variability in chemical composition and structure. Non-linear models from machine learning (ML) provide one avenue to overcoming the limitation of LFERs for single site catalysis. We select representative catalytic reaction steps, oxo formation and hydrogen-atom transfer, and analyze several strategies for incorporating ML models trained on inorganic-chemistry-appropriate, heuristic representations for relative energy predictions. We examine the promise of direct prediction to correct for scatter in weakly correlated LFERs. We also compare to baseline errors in predicting equilibrium properties of isolated complexes (spin-splitting energies, bond lengths, and redox potential) to identify how predicting catalytic properties introduces new challenges for ML model-based property prediction. We demonstrate how models can be incorporated into workflows for virtual high throughput screening of molecular oxidation catalysts. We also discuss how building heuristic models for reaction energy predictions can be implemented for multiobjective catalyst optimization and design.