Machine-Learned Electron Densities
Machine-learned electron density models predict the ground-state electron density of a structure directly, either as coefficients of atom-centered basis functions or pointwise at chosen locations in space. Because density functional theory makes the density sufficient to determine ground-state properties, a predicted density can be passed to established methods (charge partitioning, multipole calculations, or a DFT code as a starting guess) instead of training a separate model for each property.
Pointwise GNN densities for catalyst surfaces (2023)
Sunshine et al. (2023) trained a GNN on RPBE pseudo-valence densities of 32,794 OC20 structures (about 5 million grid points each) and reached a normalized density MAE of 2.54% on out-of-distribution test structures, compared with 8.30% for the superposition of atomic charge densities (SACD). Bader charges and dipole moments derived from the predicted densities were markedly better than SACD values without any training on those properties, and all errors fell with more training data. Used as an initial guess for VASP self-consistency, however, the predicted densities did not consistently converge faster than SACD, since wavefunctions and PAW augmentation occupancies also need good starting values, and a full-grid prediction still took 1-2 GPU-minutes. Summary: sunshine2023.