Hydrogen sulphide on refractory oxides: a foundation model screens configurations while DFT spectra reproduce the experimental band order
Synopsis
Whether hydrogen sulphide physisorbs or dissociates on nanoscale refractory oxide surfaces had been hard to settle because the configuration space is too large to enumerate. This work pairs unbiased automated adsorption sampling with an off-the-shelf atomistic foundation model: MACE's MPA screens candidate minima, each survivor is re-optimised with PBE-D2*(N) and HSE06 DFT, and DFT multiconfigurational infrared spectra are compared with in situ spectra at 150 K. The computed spectra reproduce the red-shift order of the H–S stretch across the three materials: silica at 2603 cm⁻¹, forsterite at 2588 cm⁻¹ and MgO at 2504 cm⁻¹.
Interpretation
Hydrogen sulphide adsorption and dissociation configurations on silica, forsterite and MgO can now be enumerated systematically without presupposing chemical intuition, yielding vibrational features comparable to experimental infrared spectra. Earlier automated adsorption workflows relied largely on grid sampling and semi-empirical methods, whose limited chemical accuracy restricted them to processes beyond chemisorption such as bond activation and dissociation. Among 16 DFT reference structures converged from 43 initial geometries on Mg2SiO4(001), MPA re-finds all 16 within a local RMSD of 0.4 Å, including all four strongly bound dissociated configurations with adsorption energies below −60 kcal mol⁻¹; UMA-OC20 and OMAT24 recover only two of those four.
The workflow first uses the foundation model to locate candidate adsorption sites in a configuration space too large for DFT, then hands every retained structure to first-principles re-optimisation, so the reported energetics and vibrational properties rest on DFT. Foundation models are trained overwhelmingly on bulk crystals, placing undercoordinated surface sites, point defects and bond-breaking events outside their fitted domain, where accuracy is not guaranteed. Across 11 matched pairs, MPA gives a mean absolute adsorption-energy error of 3.6 kcal mol⁻¹ with slope 1.04 and R² of 0.98; single-point PBE-D2*(N) calculations on MPA geometries recover the MLP energies to within 1.3 kcal mol⁻¹, while MLP-optimised structures lie on average 3.1 kcal mol⁻¹ above the DFT minima.
The differing reactivity of the three oxides appears in the computed spectra as a systematic red shift of the H–S stretch, and that order is matched in the in situ infrared spectra at 150 K. Experimentally the band maximum shifts to 2578 cm⁻¹ on Mg2SiO4 with a broad low-frequency tail and further to 2470 cm⁻¹ on MgO, while the computed maxima are 2603, 2588 and 2504 cm⁻¹ respectively. Computed spectra are equally weighted, normalised sums over the MLP-screened, DFT-refined minima and, compared with similarly normalised experimental spectra, reproduce the peak order, band shape and low-frequency tail.
Perspective
The demonstrated scope covers amorphous silica, six equilibrium-shape facets of forsterite, and MgO(100) with nearest-neighbour and next-nearest-neighbour Schottky defects, compared against in situ infrared spectra recorded at 150 K after high-temperature outgassing.
In the silica experiment the broad hydroxyl band interacting with hydrogen sulphide is centred near 3560 cm⁻¹, slightly less red-shifted than predicted, which the authors attribute to the diversity of the experimental surface versus a single-structure amorphous model; the computed MgO H–S peak is narrower than the experimental broad band because the theoretical data cover a single facet with one divacancy type. Several strongly red-shifted O–H modes on forsterite correspond to highly strained configurations that the authors expect to be quenched at operating temperatures by thermally induced atomic rearrangement.