Atomic Charge Partitioning
Atomic charge partitioning assigns a net charge to each atom by dividing space into atomic regions and summing the nuclear charge and the electron density within each one. Bader analysis, which bounds the regions by zero-flux surfaces of the density gradient, is a widely used scheme for periodic planewave calculations, and the resulting charges are used to discuss charge transfer, bonding and electrostatics.
Bader charges from predicted densities (2023)
Sunshine et al. (2023) ran Bader analysis on electron densities predicted by a graph neural network for 100 out-of-distribution OC20 test structures and compared the charges with those from the converged DFT densities. The model, never trained on charges, gave roughly an order of magnitude lower charge error than the superposition of atomic charge densities, with error decreasing as training data grew; charges obtained this way also automatically conserve total charge. Dedicated charge-prediction GNNs reported elsewhere are more accurate, but this route obtains charges as a by-product of a single density model. Summary: sunshine2023.