DFT Convergence
Plane-wave density functional theory results depend on numerical settings as well as on the functional: the k-point grid that samples the Brillouin zone, the plane-wave energy cutoff that sets the basis size, and the smearing width used to occupy states near the Fermi level, among others. A calculation is converged when tightening these settings no longer changes the quantity of interest by more than a chosen tolerance. Total energies usually converge slowly, but energy differences such as adsorption or reaction energies benefit from cancellation of errors between the terms being subtracted, so they often converge with looser settings. For large, chemically diverse datasets the settings cannot be tuned system by system, so a single conservative set is used, and whether it is adequate for every material class is an empirical question. Convergence matters for machine learning because unconverged labels can act as noise in the training data; see Neural network potentials.
Convergence of the OC20 dataset (2024)
Abdelmaqsoud and coworkers tested the OC20 settings (k-point multiplier 40, 350 eV cutoff, 0.2 eV Methfessel-Paxton smearing) on one hundred hard nonmetallic systems. Cutoff errors in total energy reached about 2 eV, and a 500 eV cutoff was needed to bring the mean absolute error near 0.03 eV; smearing errors reached about 0.15 eV, and smaller smearing widths increased the fraction of electronically unconverged runs (2.5% at 0.1 eV, 6% at 0.05 eV). On a 200k-structure subset, total energy differences at tighter settings exceeded 0.1 eV and were largest for nonmetals, yet adsorption energy differences were about 0.05 eV across all material classes because of error cancellation, and models trained at tighter settings had unchanged errors. The authors caution against assuming cancellation in general and recommend quantifying convergence on a representative subset, noting smearing scheme choices for semiconductors and Hubbard U and magnetic initialization for oxide datasets as further concerns (Abdelmaqsoud et al. 2024). Summary: abdelmaqsoud2024a. Related: DFT+U, Adsorption energy referencing.