Fig. 3: Screening high coverage NO* configurations on Pt3Sn(111).
From: Adsorbate chemical environment-based machine learning framework for heterogeneous catalysis
Configurational analysis of NO* adsorption on Pt3Sn(111), where ACE-GCN is used to predict energetics of the unrelaxed configurations generated using SurfGraph. (a) and (b) correspond to training and validation parity plots for an ACE-GCN model with NO* configurations consisting of 1–4 NO* molecules per unit cell. Test set performance and training related to configurations with 4 NO* molecules per unit cell are discussed in the Supplementary Information, Fig. S8. (c) gives predictions of the ACE-GCN model, trained on configurations of 1–4 NO* molecules per unit cell, for stability of unrelaxed 5 and 6 NO* configurations generated with SurfGraph. The predicted average BE of the unrelaxed configurations is plotted on the x-axis, while the final energy of the same configurations after DFT relaxation is plotted on the y-axis. Only configurations where the binding location of the NO* did not change after DFT relaxation are included. The ACE-GCN algorithm successfully predicts trends in adsorption energies based solely on the unrelaxed configurations generated by SurfGraph. A representative area of chemical space relevant for unstable and stable configurations is depicted on the scatter plot (c). Selected relaxed low- and high-energy configurations are shown in insets (i) and (ii), respectively. Source data are provided as a Source Data file.
