Density Matrix-Based Features as Descriptors for Oxygen Reduction and Evolution Catalysts

Developing cost-effective and efficient catalysts containing nonprecious metals is critical for chemical-to-electrical conversion technologies. The onset potentials for the fundamentally important oxygen reduction reaction, oxygen evolution reaction, and hydrogen evolution reaction can be determined from binding free energies. Herein, artificial neural networks (ANNs) were trained on a dataset of approximately 1500 metal-nitrogen-doped carbon (MNC) complexes containing first-row transition metals to predict OOH, O, OH, and H binding free energies from transferable density matrix-based features. These ANNs use density matrices from gas-phase Hartree-Fock theory with a minimal basis set for a fixed geometry to predict binding free energies at the level of solution-phase density functional theory (DFT) with a much larger basis set for optimized geometries. The ANNs were able to predict binding free energies with a mean absolute error of around 0.1 eV. Several feature selection tools such as recursive feature elimination were used to decrease the number of density matrix-based features and increase accuracy. The off-diagonal density matrix elements between the metal and ligating nitrogens were found to be especially predictive of binding free energies. This machine learning strategy has the potential to facilitate the discovery of efficient and abundant metal-based catalysts for electrochemical energy conversion.