Work function, the energy needed to remove an electron from a surface, helped predict acidic hydrogen-evolution activity on Ag-Au-Pd-Pt alloys, but it added only a small predictive lift once hydrogen adsorption was adjusted for surface coverage. The work-function-only model had a mean fit value of 0.903. The adsorption-only model scored 0.758 in its dilute 1/9 ML form, rising to 0.955 after the correction, while a combined model reached 0.969. The study asked whether work function added predictive information beyond the distribution of hydrogen adsorption energies.
The test used three combinatorial Ag-Au-Pd-Pt thin-film libraries, named Lib-1 through Lib-3, with 342 predefined measurement areas in each for characterization. The areas were screened for acidic hydrogen evolution using scanning electrochemical cell microscopy. The comparison paired those experimental readings with calculated surface descriptors.
Coverage changed the picture
To build those descriptors, the calculations used fcc(111) slabs with a 3-by-3-by-4 surface cell and nine hollow sites per slab. Each measurement area was represented by 1,000 random slabs. Across the nine sites, that produced 9,000 predicted hydrogen adsorption energies and 1,000 work-function values for each area.
The three models were jointly fitted to current densities measured at -100, -200, -300 and -400 mV versus RHE. The analysis used log-current space and a shared set of valid measurement cells. Its fit values are coefficient-of-determination scores in log-current space, so they summarize how closely each model follows the measured current pattern across the four potentials.
At 1/9 ML, the dilute setting with hydrogen represented at one of nine modeled sites, the adsorption-only mean fit value was 0.758. After correction toward a 1 ML target, it was 0.955. When the shift was fitted across the libraries, the values were 0.22, 0.20 and 0.19 eV, with a mean-field coefficient of 0.25 eV/ML. Those shifts implied an effective working coverage of roughly 0.9 ML, and the fitted-coverage combined model reached a mean fit value of 0.973. The 0.9 ML figure was an effective coverage inferred by the model, not a separate per-area coverage reading.
The signal overlapped with adsorption
Researchers also shuffled the work-function values as a control. At 1/9 ML, the combined-model score fell from 0.922 to 0.759. At 1 ML, it fell from 0.969 to 0.955. No 1/9 ML shuffle matched the unshuffled fit in 600 within-library runs or 300 leave-one-library-out runs. The result supports a work-function signal in the dilute model, while the smaller change after coverage correction points to overlap with the adsorption descriptor.
The overlap was visible directly in the descriptors. Across the three libraries, mean adsorption energy and work function had R2 values of 0.80, 0.73 and 0.51, with Pearson correlations of -0.90, -0.86 and -0.71. Pooling 1,025 measurement areas gave R2 of 0.64. Put simply, the two descriptors often moved in opposite directions, so the corrected adsorption distribution already contained much of the predictive pattern that work function could add.
The result transferred, with a proviso
To see whether the result traveled beyond a single library, the analysis used leave-one-library-out testing. Before scoring the held-out library, it recalibrated both the library-specific current scale and the fitted transport limit. Held-out fit values were 0.948, 0.879 and 0.953 for models A, B and C, respectively, with a cross-fit standard deviation of at most 0.001. This is evidence of transfer within the three-library set, but the recalibration means it is not an untouched absolute-rate prediction.
The descriptor calculations were also checked with two descriptor-specific graph neural networks on held-out compositions. The adsorption-energy network had a test mean absolute error, the average size of its prediction error, of 0.0093 eV and a fit value of 0.994. The work-function network had an error of 0.0179 eV and a fit value of 0.988. Those figures show that the surrogate models matched the study's calculated descriptor values closely on the held-out test.
A useful clue within a narrow test
That evidence has a narrow boundary. The study tested acidic hydrogen evolution on three Ag-Au-Pd-Pt libraries and modeled fcc(111) slabs, with each measurement area represented by random slabs rather than a direct map of its atomic arrangement. It therefore shows a predictive association within this selected system, not that changing work function would by itself change activity or that the same ranking will hold for other alloy spaces. The effective coverage was also inferred from a fitted correction, not directly measured for each composition.
The document is labeled a preprint, and the paper says its supporting data are openly available in Zenodo.
Paper data and sources
Original title: Work Function and High-Coverage Adsorption Energy as Hydrogen-Evolution Descriptors on Ag-Au-Pd-Pt Alloys
Authors: Zacharias Liasi, Ridha Zerdoumi, Felix Thelen et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text