volcorr

The model learned from relaxations in the Materials Project database, so it predicts the error of a calculation done the same way: PBE with PAW potentials, 520 eV cutoff, spin polarization, and a force-based ionic relaxation. Oxides and fluorides of Co, Cr, Fe, Mn, Mo, Ni, V and W get PBE+U with the Materials Project U values (Co 3.32, Cr 3.7, Fe 5.3, Mn 3.9, Mo 4.38, Ni 6.2, V 3.25, W 6.2 eV); you can override this choice above. It does not cover other functionals (for example r²SCAN) or other U values.

Only compounds whose PBE(+U) volume lies within 5 % of experiment were used for training, so materials where PBE fails badly (strongly van der Waals bound or some f-electron systems) are outside its range. Experimental volumes are at room temperature, so the prediction includes thermal expansion.

Gradient-boosted trees (LightGBM, absolute-error loss) predict y = ln(V_exp/V_DFT) from composition, shell-filling and symmetry descriptors and the calculation settings. They were trained on 29,372 Materials Project entries matched to ICSD structures; five models are averaged.

The full correction in the accompanying paper reaches 1.09 % for these calculations, but it also uses outputs of the finished DFT calculation (relaxed geometry, band gap, magnetization), which are not asked for here. A MACE-MPA-0 route needs PyTorch and is available in the command-line tool. Models: utksi/volcorr-models.