Adsorption Energy Referencing
An adsorption energy is computed from DFT total energies by subtracting the energy of the clean relaxed surface (slab) and a gas-phase reference for the adsorbate from the energy of the relaxed adsorbate+slab system. Gas-phase references are often built from linear combinations of small molecules such as H2, H2O, CO and N2. The scheme assumes that the slab under the adsorbate is essentially the same structure as the clean reference slab. When the adsorbate induces a reconstruction of the surface, the computed quantity also contains a reconstruction energy, so it is no longer a pure adsorption energy. Machine learning models trained directly on adsorption energies inherit whatever inconsistencies the referencing introduces, while models trained on total energies avoid the issue because nothing is referenced. See Neural network potentials and Coverage-dependent adsorption.
Surface reconstruction in OC20 (2024)
Abdelmaqsoud and coworkers flagged adsorbate-induced reconstruction in OC20 by comparing atom connectivity of the relaxed clean slab with that of the relaxed adsorbate+slab after removing the adsorbate. About 22% of calculations were flagged, more often for nonmetals and halides. Dropping these from validation sets, without retraining, lowered average energy MAEs of GemNet-OC, eSCN and EquiformerV2 by roughly 35% (for example 0.248 to 0.160 eV for GemNet-OC) and narrowed the gap between material classes, while force errors were unchanged since forces are not referenced. A GemNet-OC model trained on raw total energies was unaffected by reconstructions and gave adsorption energy MAEs (0.382 eV) comparable to an adsorption energy model (0.359 eV) on the IS2RE validation sets, aided by cancellation of errors between slab and adsorbate+slab predictions, and it does not need a DFT-relaxed reference slab (Abdelmaqsoud et al. 2024). Summary: abdelmaqsoud2024a. Related: DFT convergence.