Advancing the understanding and prediction accuracy of molecular adsorption energy with artificial intelligence

Molecular adsorption plays a key role in heterogeneous catalysis, with the $d$-band center theory being extensively utilized for qualitative analysis of adsorption behavior of molecules on transition-metal surfaces. However, adsorption energies predicted by $d$-band center theory can have large errors, especially when a wide range of metals is considered. The physical mechanism responsible for the deviation has been unclear, posing significant challenges of the precise design of catalysts. Here, we have integrated density-functional theory calculations with artificial intelligence approaches to develop a unifying model that quantitatively evaluates the adsorption energy of water on various transition-metal surfaces. Data mining revealed that the large deviations between the $d$-band center theory predictions and first-principles calculations primarily stem from surface relaxation effects. Additionally, we discovered that the synergistic effects of electron donation and back donation, which are mediated by the $d$ and $p$ orbitals in metals, also play a role in these deviations, although their impact is less significant compared to that of the surface relaxation. These findings enhance our understanding of molecular adsorption behavior and are poised to influence research fields such as surface science and heterogeneous catalysis.