MACE-MP misses literature values for graphene on Ni(111), Mat3ra finds

The notebook addresses a practical question for materials modeling: when is a fast machine-learning force field good enough, and when does it break down? Graphene on Ni(111) looks simple but its binding depends on where carbon atoms sit relative to the nickel surface, across four registries—fcc, hcp, hollow, and bridge. The authors note that bridge was not quantified in the reference papers they used, so they added it as an extra point.

The workflow is split into two tiers. Tier 1 uses MACE-MP, a machine-learning force field, to relax each registry in minutes, yielding an overview of energies and initial structures. Tier 2 runs DFT jobs with the LDA functional, spin polarization, fixed-cell relaxation, and no dispersion correction, then compares relaxed structures and work of adhesion against literature values. The key finding is that MACE-MP, trained on PBE data, does not reproduce literature values for this interface: chemisorption is underestimated by several times and distances are shorter. The notebook explicitly labels Tier 1 as a screening tool, not an answer.

The stakes here are about trust in fast surrogates. For researchers screening many metal/2D interfaces, MACE-MP can quickly flag candidates, but the Ni(111)/graphene case shows the gap can be large enough to change conclusions. The outcome hinges on whether PBE-trained models can be corrected for this chemistry, and on whether the same discrepancy appears for other metals or 2D materials—something the notebook's template is designed to test.