Reference Conventions and Structure Types, Rather Than 4f Occupancy, Shape Lanthanide Errors in Universal Machine-Learning Interatomic Potentials

Compounds that pair a lanthanide with a d-block metal and a p-block anion sit behind permanent magnets, catalysts and energy-storage materials, yet large parts of this chemical space have not been computed. Universal machine-learning interatomic potentials now make it possible to screen such spaces for stability at a small fraction of the cost of density functional theory, and they are increasingly used to decide which candidates are worth calculating properly. Their training data, however, hold few lanthanide compounds, describe the same element with different pseudopotentials from one database to the next, and carry empirical energy corrections fitted on compound sets from which lanthanides were excluded, so their reliability for felectron chemistry has not been established. Here we benchmark five universal potentials against density functional theory computed with Materials Project settings on 82 lanthanide–d-block–anion ternaries covering the lanthanide series (excluding Pm) and seven anions. Apparent accuracy turns out to be set mainly by the energy convention of the reference: judged against the quantity each model was trained to predict, all five reach mean absolute errors of 0.07 to 0.10 eV per atom, while the wrong convention inflates the error by more than a factor of three. Error does not follow 4f occupancy or the 4f pseudopotential treatment assigned to each element. It is larger for intermetallic than for ionic compounds and concentrates in two structure types, the CeNiC2-type carbides and the PbFCl-type phosphides, which most models miss even though these compounds have Materials Project entries. For ytterbium, every model follows the divalent convention of the older Materials Project data rather than the trivalent one now in use, a singleelement inconsistency of roughly 0.2 to 0.3 eV per atom. These results indicate where lanthanide chemical space can be screened with current potentials and where explicit DFT validation is still required, and they point to structural and data-provenance diagnostics, rather than elemental ones, for the next generation of models.