Content

• SCF convergence problems
  • parameters recommendations
    • CONTROL namecard
    • SYSTEM namecard
    • ELECTRONS namecard
    • K_POINTS namecard
  • too many eigenvalues not converged found in the last iteration steps
  • can not reach convergence till all scf steps
  • too many processors
  • fatal error in internal_gather: message truncated, error stack:
• Geometry
  • If band gap exists in ground state calculation: need test and example

SCF convergence problems

parameters recommendations

CONTROL namecard

• nstep = 1000
• forc_conv_thr = 1e-5 in Ry/Bohr, sometimes can not converge, maybe a little bit too high for some cases, equivalent to 2.57e-4 eV/Angstr
• etot_conv_thr = 1e-6 always good if no SCF convergence problems
• dipfield = .true., tefield = .true. for using dipole correction for polar slab system with vacuum layer

  SYSTEM namecard

• ecutwfc see recommended value from each pseudopotential files \(\Delta E < 2\times10^{-5} \rm Ry/atom\)
• ecutrho can be 8-10ecutwfc for Ultrasoft, usually 4ecutwfc (default) is ok
• nosym = .true., force_symmorphic = .true. for defected, slab, not bulk system, the latter is for yambo
• occupations = 'smearing', smearing = 'mp', degauss = 1e-3 about 300 K, for conductor should do convergence test of degauss, for insulator can choose occupations = 'fixed' (also need for Berry phase calculation)
• degauss smearing contribution as small as possible \(<1\times10^{-4} \rm Ry/atom\)
• edir = 3, eamp = 0.0, emaxpos = 0.9, eopreg = 0.02 for vacuum layer at z-direction, slab located at the middle of unit cell, emap means balanced charged sides of slab, emaxpos refers to the location where to appear correction, eopreg refers to the length of correction
• edir = 3, eamp = 0.0, emaxpos = 0.8, eopreg = 0.02 for slab located at bottom of unit cell, check carefully if emaxpos go across the slab model

  ELECTRONS namecard

• electron_maxstep = 500 good
• mixing_mode = 'plain' for common cases, mixing_mode = 'local-TF' for reduced-dimensional cases
• mixing_beta = 0.6 for common cases, mixing_beta = 0.3 for 2D cases
• mixing_ndim = 12 for local PC, mixing_ndim = 16 for HPC first try
• conv_thr = 1e-11 is ok for scf and nscf, conv_thr = 1e-9 for relax and vc-relax
• diagonalization = 'david' for common cases, diagonalization = 'cg' much slower but robust, can try if not converged.

During nscf (as well as bands) KS eigenvalues are diagonal to full accuracy (diago_full_acc=.true. by default), at variance with scf calculations, where accuracy is lowered for empty states (not relevant to compute the density not the total energy). While one could set diago_full_acc=.true. in scf runs, it is anyway highly recommended to run a second nscf calculation to perform BZ refinements or increase the number of state for plotting or inspection.

K_POINTS namecard

8 8 2 0 0 0 no grid shift if doing yambo

too many eigenvalues not converged found in the last iteration steps

• Changing pseudopotential types: USPP, NCPP, PAW PP
• Increase ecutwfc with ecutrho=4*ecutwfc (for ONCVP, 8-10 for USPP)
• Decreasing conv_thr parameter (scf 1e-11, relax 1e-9)
• Decreasing mixing_beta (0.1-0.3 is ok for slab)
• Changing ion_dynamics and cell_dynamics (usually bfgs)
• Decreasing k meshing

can not reach convergence till all scf steps

• reduce mixing_beta default 0.7, can be 0.1-0.3 for slab model (<0.05 for very long slab). It reudces the mixing ratio of charge density from the previous iterative steps
• increase mixing_ndim to 12 (for local PC) or higher (16 for HPC is totally ok and can be more larger like 24). Determine how many iterative steps are considered for charge density mixing of mixing_beta
• use mixing_mode of local-TF for slab model
• change

too many processors

No share memory parallel (OpenMP) in my computer

OMP_NUM_THREADS=1

in batch file (or ~/.bashrc file for local computer)

fatal error in internal_gather: message truncated, error stack:

Similar things:

• Caught signal 11 (Segmentation fault: address not mapped to object at address (nil))
• message…

Problems relate to MPI, we call mpi out of memory, may be cause by the long address and inter-overlapping

can try:

• use fewer cores (just test in local PC first and see whether it can start calculate)
• incresae the allocation of memory (in slurm, that is #SBATCH --mem=600G), useful for very large system

Geometry

Provided known space_group, don’t use ibrav ATOMIC_POSITION crystal_sg nat is supposed to accord to inequivalent atoms, which is, irreducible form of atoms in primitive cell Irreducible form of atoms can be obtained from VESTA by loading primitive cell file

If band gap exists in ground state calculation: need test and example

Use occupation = 'fixed', tot_magnetization with accordance to previous SCF results, nbnd with accordance to number of Kohn-Sham states in previous SCF results. The use of startingpot and startingwfc (in &Electron) is recommended to reduce CPU time and avoid from divergence