CO Adsorption Puzzle in Surface Science
- CO adsorption puzzle is defined by discrepancies between experimental observations and theoretical predictions of CO binding on metal and oxide surfaces.
- Studies show that standard DFT methods mispredict adsorption sites due to density-driven and self-interaction errors, prompting the use of hybrid, DFT+U, and many-body techniques.
- Advances in machine learning and improved computational methods now enable more accurate predictions, aligning theoretical models with experimental data for catalyst design.
The CO adsorption puzzle refers to persistent discrepancies between theoretical predictions and experimental observations of carbon monoxide (CO) adsorption site preferences and energetics on metal and oxide surfaces. Classical models and conventional density functional theory (DFT) methods frequently fail to predict the correct binding site (particularly for late transition metals such as Pt, Rh, and Cu) and often misrepresent coverage-dependent effects, phase transitions, or the interplay of surface defects and adsorbate interactions. Multiple theoretical, experimental, and computational studies have illuminated various facets of this problem, leading to significant revisions of models and methodologies in surface science and heterogeneous catalysis.
1. Theoretical Origins of the Puzzle
At the core of the CO adsorption puzzle is the persistent failure of commonly used generalized gradient approximation (GGA) DFT functionals to predict experimentally observed top-site adsorption of CO at low coverage on late transition-metal surfaces (e.g., Pt(111), Rh(111), Cu(111)). Instead, GGA functionals overstabilize high-coordination binding sites like fcc and hcp hollows. Analysis has shown that this error primarily arises from self-interaction and density-driven errors that incorrectly position the molecular orbitals of CO relative to the Fermi level of the substrate, leading to excessive back-donation from the metal surface to the CO 2π* orbital. The erroneous site preference is not rectified by more advanced semilocal approaches such as meta-GGA (SCAN) and is only partially alleviated by hybrid functionals like HSE06, whose computational expense limits their practical applicability for large or complex systems (Patra et al., 2018, Liang et al., 29 Jul 2025).
2. Model System Studies: ZGB and Oxygen Adsorption Mechanisms
The Ziff–Gulari–Barshad (ZGB) lattice-gas model, a kinetic Monte Carlo framework for catalytic CO oxidation, historically exhibited a continuous transition between an oxygen-poisoned state and the reactive regime that was not seen in experiments. The standard ZGB model's O₂ adsorption step places two O atoms on nearest-neighbor sites, which induces unphysical oxygen poisoning and a sharp phase boundary at critical CO partial pressure. Experimental STM results indicating "hot atom" recoils upon O₂ dissociation inspired the introduction of modified oxygen entrance rules: spatially separated O atoms (either at next-nearest neighbors or specific pairs). Three variants—ZGB-n, Ha1, and Ha2—implement such constraints, effectively eliminating the nonphysical continuous transition. The revised phase diagrams, especially when augmented with a CO desorption mechanism to mimic temperature effects, align with experimental findings by enabling immediate onset of CO₂ production as soon as CO partial pressure is nonzero (1112.1433).
| Model | O₂ Dissociation Rule | Phase Transition at y₁ |
|---|---|---|
| Standard ZGB | Both O atoms on nn sites | Present (unphysical) |
| ZGB-n | O atoms at target + nnn site (√2 apart) | Eliminated |
| Ha1/Ha2 | O atoms only on selected empty vertical or horizontal pairs, or with relaxed center occupation | Eliminated |
3. Experimental Insights from Oxides and Alloys
Extensive STM, XPS, TPD, and complementary DFT studies on oxide and alloy surfaces have further elucidated the complexities of the CO adsorption process:
- On Fe₃O₄(001), CO binding is weak and physisorptive at regular sites (activation energy ≈0.28 eV) but is stronger at specific defect sites. A coverage-driven transition from 2D mobile gas to an ordered overlayer is observed, with enthalpy-entropy compensation dictating adsorption energetics and kinetics. No evidence is found for surface reduction or carburization by CO, underscoring the importance of defect states and collective interactions (Hulva et al., 2018).
- On Sr₃Ru₂O₇(001) and Ca₃Ru₂O₇(001), CO first forms a weakly bound precursor at apical surface O (binding energy around –0.7 to –0.85 eV), but upon thermal or STM-induced activation, CO replaces the apical O, forming a metal-carboxylate complex (Ru–COO) with much stronger binding (up to –2.2 eV), accompanied by local octahedra distortion and increased surface mobility. The process provides a mechanistic basis for the "UHV aging effect" in these perovskite oxides (Stöger et al., 2018, Mayr-Schmölzer et al., 2018).
- On TiO₂(110), polaron formation and the distribution of oxygen vacancies modulate CO binding, with CO–polaron complexes forming distinct signatures in STM and tunable adsorption energetics depending on the reduction state and spatial coupling (Reticcioli et al., 2018).
- On doped or defective 2D semiconductors (e.g., GaN-ML), pristine surfaces exhibit weak physisorption, but N-vacancy defected monolayers and certain dopants (e.g., Fe) display dramatically enhanced binding, pronounced charge transfer, and even electronic and magnetic transitions, which are key for sensing applications (Li et al., 2023).
4. Computational Solutions and Methodological Advances
Efforts to resolve the CO adsorption puzzle on transition-metal surfaces have led to several computational innovations:
- Density-driven errors, particularly in the context of self-consistent GGA or meta-GGA functionals, are now recognized as a major source of incorrect site preference and overbinding. Non-self-consistent approaches, such as PBE evaluated on densities generated with PBE+U applied to C and O atomic p orbitals, have been shown to recover experiment-consistent top-site adsorption for CO/Pt(111). This correction primarily adjusts the HOMO–LUMO gap of CO, reduces artificial back-donation, and restores energetic ordering (Patra et al., 2018).
- Many-body treatments such as the Random Phase Approximation (RPA) resolve the surface and adsorption energetics as well as site preferences, placing the top site correctly and matching adsorption energies with experiment. The prohibitive computational cost of RPA has motivated the use of Δ-machine learning (Δ-ML) approaches and machine-learned force fields (MLFFs) that are trained on representative subset corrections (PBE→RPA), providing near-RPA accuracy for coverage-dependent adsorption with greatly improved efficiency (Liu et al., 2022).
- Machine-learned exchange-correlation functionals (DeePKS), trained to match hybrid functional (HSE06) results, achieve the correct adsorption site preferences, PES, and structural parameters for CO on Cu(111) and Rh(111), with site energy differences within ≈10 meV of reference values. Such models offer scalability and transferability critical for high-throughput screening of catalytic materials (Liang et al., 29 Jul 2025).
| Approach | Key Feature | Success on CO/Solid Problem |
|---|---|---|
| GGA (PBE) | Semilocal; prone to self-interaction | Overbinds, wrong site |
| Hybrid (HSE06) | Mixes exact exchange | Corrects site, costly |
| DFT+U | Adds U to CO p-orbitals (density correction) | Recovers top site on Pt(111) |
| RPA/MLFF | Many-body, Δ-learning w/ML correction | Site preference, coverage |
| DeePKS | Neural network XC functional, hybrid-trained | Site, PES, efficient |
5. Coverage Effects, Ensemble/Ligand Influences, and Alloying
The energetic and structural properties of CO adsorption are strongly affected by surface coverage, alloy composition, and ensemble size:
- On metal surfaces and supported clusters, ligand effects (electronic modification due to neighbors, e.g., Pd by Au) and ensemble effects (site size and geometry) both modulate adsorption energy, but ensemble effects are dominant. For PdAu clusters, as Au segregates to the surface and Pd site continuity breaks, adsorption site preference migrates from hollow to bridge to atop, with a marked decrease in adsorption energy—up to 36–70% upon increasing Au content (Sitja et al., 2020).
- Machine-learned potentials confirm that at low coverage, CO–CO interactions are minimal and top-site adsorption prevails, but with increasing coverage, a sequence of ordered motifs emerges, eventually leading to a saturation near 13/16 ML on Rh(111), a result consistent with experimental STM patterns (Liu et al., 2022).
6. Surface Chemistry Beyond Metals: Defects, Dynamics, and Complex Materials
- On MoSe₂ with Se vacancy and O adatoms, static DFT indicates exothermic recombination of CO and O_top, but AIMD simulations reveal that even above-threshold incident energies, CO₂ formation is vanishingly rare. Entrance-channel barriers, orientation, and dynamical constraints dominate over thermodynamic driving force, underscoring the importance of reaction dynamics for surface oxidation chemistry (Bombín et al., 22 Sep 2024).
- On MgO(001), periodic CCSD(T) calculations yield adsorption energies and vibrational frequency shifts for CO that are in excellent agreement with experiments. Many DFT functionals, by contrast, overbind and predict even the wrong sign of frequency shift due to self-interaction errors, with these deficiencies only partially ameliorated by hybrid functional corrections. The success of local-correlation wavefunction methods demonstrates their value for benchmarking and driving improvements in surface modeling (Ye et al., 2023).
7. Implications for Catalysis and Surface Science
Resolution of the CO adsorption puzzle has significant consequences for catalyst discovery, the interpretation of spectroscopic data, and the rational design of surface materials. Accurate modeling of adsorption geometry and energetics is central to understanding reaction mechanisms, phase stability, and dynamic response under catalytic conditions. Advanced machine-learning functionals and force fields, density correction strategies, and explicit consideration of surface structure, defect chemistry, and ensemble/ligand effects are requisite for predictive surface science and computational catalyst design. Robust benchmarking against high-level ab initio (CCSD(T), RPA) and careful integration of machine learning methodologies are central to progress in this area.
In conclusion, the CO adsorption puzzle highlights the critical interplay between electronic structure theory, many-body and machine-learning approaches, surface morphology, and dynamical effects in the accurate description of molecule–surface interactions. Its ongoing resolution continues to drive both methodological innovation and the fundamental understanding of surface reactivity in heterogeneous catalysis.