SCF convergence threshold on the charge-density residual (ABACUS scf_thr; not an energy). The default is ABACUS's own LCAO default and what forces and stress need: at 1e-6 a relaxation already ends at a different volume. Loosen to 1e-4 only for quick screening of energies. Force and stress checks (relax, MAE's relaxed-geometry check) always run at 1e-7 or tighter.

Range: 0.05 to 1

K-point spacing in 1/Å. The default is finer than the other routes' 0.3. This is the mesh of the self-consistent run; the spin–orbit energy is summed on sum_kspacing.

Turn on DFT+U with an effective U in eV per element, e.g. {"Ni": 6.2}. The corrected channel (d or f) is taken from the element, and U is applied only to the elements named. Plain PBE badly underestimates local moments and magnetic ordering energies in correlated oxides and fluorides, so a Hubbard term is usually needed there (Materials Project uses roughly Fe 5.3, Co 3.32, Ni 6.2, Mn 3.9, Cr 3.7, V 3.25, Cu 4.0). Setting this also softens the default mixing_beta to 0.2 and allows more SCF iterations, which correlated oxides need; an explicit mixing_beta still wins. Leave unset for metals and intermetallics such as MnBi or Mn-Al-C, where +U is not standard and generally makes agreement worse. The scheme is Dudarev, so this is U minus Hund J, not bare U.

LCAO basis size: SZ (fastest), DZP (balanced), TZDP (most accurate)

Magnetization directions to test, named in the crystal's frame: 001 is the unique axis of a tetragonal, hexagonal or trigonal crystal and 100 and 010 lie in the plane normal to it. A cell whose own axes already are the symmetry axes is read as written. Any other cell, such as a primitive body-centered one, is read in the frame of its conventional cell (c along z, a along x). Monoclinic and triclinic cells keep their own Cartesian axes. The response's direction_frame gives the frame used and each direction as a Cartesian vector.

Range: to 1

Charge mixing step (0–1). Left unset, it is 0.2 for structures with magnetic elements (Fe, Co, Ni, Mn, Cr, rare earths) and 0.4 otherwise. An SCF that does not converge is continued once from its own density at half this value before the run fails; if it still fails, try a quarter (0.05 for a magnet). Frustrated Mn compounds whose moment keeps changing between iterations need a different initial_magmoms seed rather than smaller mixing.

Range: 0.03 to 1

K-point spacing in 1/Å for the sum of the spin–orbit energy over the Brillouin zone, which sets the anisotropy and converges far more slowly than the self-consistent run. It reuses that run's Hamiltonian, so a finer value costs no extra DFT. Never coarser than kspacing. Lower it to 0.08 to check a result that matters, above all with W or Ta.

XC functional

If false (default), reject CIFs whose DFT max force or stress exceeds the relax thresholds so MAE is for a DFT energy minimum. Run /dft/structure/relax first, or set true to evaluate the uploaded geometry as-is.

Signed starting moments in µB, one per atom in CIF site order. Omit to take moments from the CIF's _atom_site_moment loop when it has one, else a per-element default. Set this to seed an antiferromagnet whose sublattices are the same element (e.g. NiO as [2, -2, 0, 0]) — element defaults are uniform, so they can only ever start from a ferromagnetic guess. Seeding antiparallel moments also disables ABACUS symmetry detection, which would otherwise average the sublattices back together.

Range: to 1

Occupation smearing width in eV (converted to Rydberg for ABACUS). Typical metals: 0.05–0.10 eV. Gaps need ~0.05 eV or smaller.

Evaluate the primitive cell instead of the cell as uploaded. Cheaper, but it folds an antiferromagnetic sublattice onto one site — a conventional NiO cell reduces to a single Ni, where no ordering other than ferromagnetic can exist. Leave false for any magnetic ordering question.