Unifying theory of electronic descriptors of metal surfaces upon perturbation
Yang Huang, Shih-Han Wang, Mohith Kamanuru, Luke E. K. Achenie, John R. Kitchin, and Hongliang Xin, “Unifying theory of electronic descriptors of metal surfaces upon perturbation”, Physical Review B 110(12), L121404 (2024). https://doi.org/10.1103/physrevb.110.l121404
This summary was written by a language model from the paper and has not been checked line by line. Rely on the published paper for details.
Question
The d-band center of a surface metal atom is a standard descriptor of its reactivity, and its shift when the site is perturbed (by alloying, strain or a change of facet) is conventionally explained by strain, ligand and charge transfer effects. These act together and are hard to separate, while purely data-driven machine learning models can predict the d-band center but give little physical insight. The Letter asks whether a deep learning model with physics built into it can predict the d-band center of transition and noble metal sites accurately and also decompose the shift into its physical causes, including any that the conventional picture misses (Huang et al. 2024).
Approach
The starting point is a rectangular d-band model that writes the d-band center of a site in terms of its d-band width and fractional d-band filling, relative to a reference atom in the bulk fcc metal. Holding the filling fixed, a wider band sits lower in energy; a second term accounts for any change in filling from charge transfer. The width is taken proportional to the square root of the second moment of the d band, which is computed from tight-binding hopping integrals between the site’s d orbitals and the orbitals of all neighbors within a 5.5 Å cutoff. These hopping integrals depend on the d-orbital radii (the ligand effect) and on interatomic distances (the strain effect), following the solid-state table of Harrison (Huang et al. 2024).
The model is then parameterized with learned corrections: an atom-specific shape factor alpha that corrects the rectangular-band assumption, an atom-specific coefficient beta multiplying a local Mulliken electronegativity difference to capture charge transfer, a bond-wise factor zeta that rescales the unrelaxed interatomic distances, and an element-wise factor lambda on the d-orbital radius. A graph convolutional network predicts these variables, and the physics module turns them into the d-band center, filling and width, all of which enter the loss against DFT values. This is the theory-infused neural network (TinNet) framework. After training, SHAP (Shapley additive explanations) analysis attributes the d-band center shift relative to the pure fcc metal of the same site type to four variable pairs: charge transfer, strain, ligand and the shape factor (Huang et al. 2024).
Key findings
The dataset contains 12,165 unique surface sites on fcc surfaces with varied composition, structure, facet and external strain. Under nested 10-fold cross-validation, TinNet reaches a mean absolute error of 0.032 eV for the d-band center, similar to a plain graph convolutional network, while a fully connected network on simple tight-binding features gives about 0.13 eV and the purely physical rectangular-band model with d-charge conservation stays near 0.27 eV (Huang et al. 2024).
The learned charge transfer term is close to zero across the data, consistent with d-charge conservation, and its mean absolute Shapley value is only 0.01 eV. Strain and ligand effects each contribute about 0.15 eV. The largest contribution, about 0.20 eV, comes from the shape factor alpha, which has a broad distribution around one and which the authors attribute to a previously overlooked resonance effect: energy misalignment between interacting orbitals that changes the band shape (skewness, kurtosis) without changing its second moment. The bond factor zeta clusters around 1.0 with a standard deviation of 0.04 and tracks DFT-relaxed bond lengths for strained Pd(111) (+5% to -5%) and PdAg alloys, with systems under tensile strain tending to contract on relaxation (Huang et al. 2024).
For Pt monolayers on fcc(111) substrates, strain dominates the shift. The Shapley values for charge transfer, strain and ligand effects correlate linearly with substrate electronegativity, bulk lattice constant and d-orbital radius, while the resonance contribution correlates strongly with the bulk d-band center of the substrate. Among 5d substrates, Pt on Re has the largest d-band skewness (+0.48), as unoccupied states skewed to high energy push the distribution down to preserve the filling (Huang et al. 2024).
Significance
The work extends the constant-filling, matrix-element picture of strain and ligand effects into a learned model that keeps the same physical variables, adds a fourth mechanism (orbital resonance acting through band shape), and quantifies the relative size of all four over a large dataset. It shows that physics-embedded deep learning can match black-box accuracy while remaining interpretable, which the authors present as a route to theory-guided rather than brute-force catalyst design. Data and notebooks are released on GitHub (hlxin/tinnetdos) (Huang et al. 2024).