RESUMO
Predicting adsorption energies of small molecules (e.g., OH, OOH, CO) on electrocatalysts involved in electrochemical reactions aids in accelerating the design and screening of electrocatalysts. Avoiding computationally expensive electronic structure calculations increases the speed of such predictions. Geometric and electronic descriptors have been reported to characterize the environment around surface active sites and predict adsorption energies. However, these descriptors cannot be used to predict adsorption energies of small molecules on various substrates, e.g., metal-oxide and nonmetal electrocatalysts. We compare the performance of these descriptors in predicting adsorption energies of small molecules on various electrocatalysts with adsorption energies calculated from density functional theory. We show that two recently developed machine learning algorithms, Crystal Graph Convolutional Neural Network (CGCNN) and Atomistic Line Graph Neural Network (ALIGNN), outperform the reported descriptors based on geometric (coordination number of the active site and its nearest neighbors) and electronic (the bond-energy-integrated orbitalwise coordination number, the electronegativity, and the number of valence electrons of the active site) properties in predicting the adsorption energies. Our results suggest that ALIGNN is almost always more accurate than CGCNN in adsorption energy predictions. The improvement ranges from 0.02 to 1.0 eV in the mean absolute errors (MAEs). We also compare the performance of CGCNN and ALIGNN algorithms in predicting the overpotentials of the oxygen evolution reaction occurring on various electrocatalysts with MAEs of 0.06 and 0.05 V, respectively.
