Fig. 3: Enhancing prediction accuracy by distinguishing similar adsorption motifs.

a Parity plots of the ML-predicted (GAT-w/oCN) versus DFT-calculated formation energy of M-C on the 3-fold-only Cads database. Mean absolute error (MAE) and R2 values are provided in the parity plot; the violin plot in the inset shows the absolute error distributions, and the inner dash line represents the median (unit: eV). The pairs of hcp-/fcc-hollow site adsorption motifs in the weak adsorption range (b) with identical ML predictions derived from the GAT-w/oCN joined with dashed lines. c Results colored with the differences in the ML predictions obtained from the GAT-wCN; the diagram in the inset shows the predictions from GAT-w/oCN and GAT-wCN plotted with the hexagons and circles, respectively. The 3-fold-only Cads Database with 3262 entries is provided in GitHub repository at Data Availability. ML machine learning, DFT density functional theory, GAT graph attention network, CN coordination number.