Direct spin-difference initialization for spin-polarized PAW DFT

Develop a machine-learning method that directly initializes the smooth spin-difference density and its corresponding spin-difference PAW augmentation occupancies for spin-polarized projector augmented-wave density functional theory calculations.

Background

Spin-polarized PAW calculations require an initial magnetic state. This can be supplied either through atomic magnetic moments, which a DFT code expands into a real-space initial magnetization, or by directly providing the smooth spin-difference density together with its spin-difference PAW augmentation occupancies. The paper develops machine-learning approaches for both types of initialization but identifies direct prediction of the spin-dependent density components as an unresolved challenge and does not address it within the stated scope. This problem is therefore the explicit unresolved extension needed for fully direct, reference-free spin initialization in PAW DFT.

References

While both can be modeled with machine learning, they pose significant challenges, respectively, with the latter remaining an open challenge in the field and outside the scope of this work.

— Complete Neural Electronic Initialization Accelerates Materials DFT  (2609.21759 - Ærtebjerg et al., 18 Sep 2026) in Appendix B.2, “Magnetic Moment Initialization”