Complete Neural Charge Density Initialization for End-to-End Acceleration of Materials DFT

Density functional theory (DFT) consumes substantial high-performance computing resources. In materials science, plane-wave DFT is the dominant framework, and machine learning models could accelerate calculations by providing improved initial guesses at electronic densities and parameters. However, existing evaluations reuse PAW augmentation and spin components from converged DFT calculations, meaning that true cold-start acceleration has not yet been demonstrated using machine learning models. We show that these missing components affect SCF convergence and must be treated explicitly. We introduce AugNet, the first general equivariant model for predicting PAW augmentation occupancies, and combine it with separate state-of-the-art models for valence density and magnetic moment prediction. Together, these models provide all structure-dependent electronic initialization components required for general spin-polarized PAW calculations directly from atomic structure. The resulting pipeline achieves up to a $\sim26$ % reduction in plane-wave DFT wall time on unseen structures without relying on any electronic quantities from converged calculations.