Investigating CO adsorption on Cu(111) and Rh(111) surfaces using machine learning exchange-correlation functionals

Abstract The ‘CO adsorption puzzle’, a persistent failure of utilizing generalized gradient approximations in density functional theory to replicate CO’s experimental preference for top-site adsorption on transition-metal surfaces, remains a critical barrier in surface chemistry. While hybrid functionals such as HSE06 partially resolve this discrepancy, their prohibitive computational cost limits broader applications. We tackle this issue by adopting the Deep Kohn-Sham (DeePKS) method to train machine-learned exchange-correlation functionals. Principal component analysis reveals that the input descriptors for electronic structures separate distinctly across different chemical environments, enabling the DeePKS models to generalize to multi-element systems. We train system-specific DeePKS models for transition-metal surfaces Cu(111) and Rh(111). These models successfully recover experimental site preferences, yielding adsorption energy differences of about 10 meV compared to HSE06, and can be applied to study unseen adsorbate coverages. Furthermore, a single model for the two surfaces is trained, and the model achieves comparable accuracy in predicting not only adsorption energies and site preference but also potential energy surfaces and relaxed surface adsorption structures. The work demonstrates a promising path towards universal models, enabling catalyst exploration with hybrid functional accuracy at substantially reduced cost.