Nonconverging electronic minimization for specific magnetic systems
Posted: Sun May 10, 2009 6:16 pm
Dear VASP users and developers,
I decided to re-post the description of my problem (also posted at http://cms.mpi.univie.ac.at/vasp-forum/ ... php?4.5586) to this forum. Meanwhile I have discovered another several structures with similar convergence problems, i.e. it seems that the scf-problem is systematic. Meanwhile one of the "bad" systems was tested with our home-made code, and it was possible to achieve the desired accuracy. Although the test was done employing norm-conserving pseudopotentials and not with the paw, it suggests that the problem resides on the minimizer, rather than on the system. Please find the detailed description of the problem below.
I experience severe convergence problem with relatively complex magnetic structure (pure iron, 32 atoms cubic fcc-cell, special distribution of 16 up and 16 down magnetic moments, PAW PBE potentials without 3p-states, NUPDOWN=0, initial atomic positions are nonrelaxed fcc-sites). During electronic minimization of the supercell with local magnetic moments smaller than ~1.5-1.7 Bohr magneton (depends on the supercell volume) the total energy is fluctuating with the amplitude of ~10^-4 Hartree/cell, and is not converging within 50 steps. Inspection of the total energy reveals that the energy can even smoothly increase during several (~10) steps, i.e. it does not show stable tendency to decrease. By taking snapshots of the local magnetic moments one recognizes that some of the moments undergo the sign reversal during the minimization.
Since such a nonconverging energy indicates that the residuum density does not vanish upon minimization I have tried to stabilize the electronic minimizer. First I have reduced AMIX_MAG to 0.1 to reduce spin fluctuations, and used BMIX*=0.00001 to keep the speed of the convergence high. This stabilizes magnetic moments (removes their flips), but does not improve the convergence. Closer check of the magnetic moment vs iteration shows that, although most of the moments do not significantly change, they are still “breathing� (e.g., some of the moments after smooth decrease during 60 iterations start to fluctuate with amplitude ~0.1 Bohr magneton) and lead to change of the total energy.
To decrease this “breathing� I have played with BMIX and BMIX_MAG by changing them from 0.0000001 to 1.0, without any visible improvement of the convergence.
Furthermore, I have tried several other options in different combinations:
(*) included LMAXMIX=6
(*) checked that k-points mesh and energy cut-off are not responsible for such ill behavior
(*) checked Pulay mixer with different settings of AMIX* and BMIX*
(*) changed the memory of the Broyden mixer with MAXMIX from 7 to 45
(*) checked ALGO=Normal
(*) tried to increase NELMDL up to 15
(*) tried calculations removing NUPDOWN=0 tag
None of the setups was successful, i.e. the energy was fluctuating and its convergence was not better than ~ 10^-4 Hartree/cell even when 100 steps were used.
At the same time, the default settings of the INCAR at larger volumes (and, consequently, larger magnetic moments) provide a reasonable electronic convergence without flips of the local magnetic moments. In fact, the moments vs iteration curves are converged rather fast and, therefore, do not destabilize the minimization routine. Also no problems occur at small volumes where magnetic moments disappear.
I would like to ask whether someone else has experienced such nonconverging magnetic systems, and whether there exists a known solution of this issue. Any advise how to overcome the problem will be, of course, also highly appreciated. Thank you very much for your response.
With best regards,
Alexey Dick
P.S. Parallel/serial VASP v.4.6.31 compiled with Intel compiler was used.
The typical INCAR file that I have used is provided below:
PREC = accurate
ISMEAR = 1
SIGMA = 0.15
ENCUT = 270
ISPIN = 2
MAGMOM = 16*5.0 16*-5.0
ALGO = NORMAL
EDIFF = 1.0E-06
NELM = 100
NELMIN = 7
NELMDL = -7
ADDGRID = .TRUE.
ISTART = 0
ICHARG = 2
NBANDS = 200
NUPDOWN = 0
LMAXMIX = 6
AMIX = 0.2
AMIX_MAG = 0.1
BMIX = 0.0001
BMIX_MAG = 0.0001
I decided to re-post the description of my problem (also posted at http://cms.mpi.univie.ac.at/vasp-forum/ ... php?4.5586) to this forum. Meanwhile I have discovered another several structures with similar convergence problems, i.e. it seems that the scf-problem is systematic. Meanwhile one of the "bad" systems was tested with our home-made code, and it was possible to achieve the desired accuracy. Although the test was done employing norm-conserving pseudopotentials and not with the paw, it suggests that the problem resides on the minimizer, rather than on the system. Please find the detailed description of the problem below.
I experience severe convergence problem with relatively complex magnetic structure (pure iron, 32 atoms cubic fcc-cell, special distribution of 16 up and 16 down magnetic moments, PAW PBE potentials without 3p-states, NUPDOWN=0, initial atomic positions are nonrelaxed fcc-sites). During electronic minimization of the supercell with local magnetic moments smaller than ~1.5-1.7 Bohr magneton (depends on the supercell volume) the total energy is fluctuating with the amplitude of ~10^-4 Hartree/cell, and is not converging within 50 steps. Inspection of the total energy reveals that the energy can even smoothly increase during several (~10) steps, i.e. it does not show stable tendency to decrease. By taking snapshots of the local magnetic moments one recognizes that some of the moments undergo the sign reversal during the minimization.
Since such a nonconverging energy indicates that the residuum density does not vanish upon minimization I have tried to stabilize the electronic minimizer. First I have reduced AMIX_MAG to 0.1 to reduce spin fluctuations, and used BMIX*=0.00001 to keep the speed of the convergence high. This stabilizes magnetic moments (removes their flips), but does not improve the convergence. Closer check of the magnetic moment vs iteration shows that, although most of the moments do not significantly change, they are still “breathing� (e.g., some of the moments after smooth decrease during 60 iterations start to fluctuate with amplitude ~0.1 Bohr magneton) and lead to change of the total energy.
To decrease this “breathing� I have played with BMIX and BMIX_MAG by changing them from 0.0000001 to 1.0, without any visible improvement of the convergence.
Furthermore, I have tried several other options in different combinations:
(*) included LMAXMIX=6
(*) checked that k-points mesh and energy cut-off are not responsible for such ill behavior
(*) checked Pulay mixer with different settings of AMIX* and BMIX*
(*) changed the memory of the Broyden mixer with MAXMIX from 7 to 45
(*) checked ALGO=Normal
(*) tried to increase NELMDL up to 15
(*) tried calculations removing NUPDOWN=0 tag
None of the setups was successful, i.e. the energy was fluctuating and its convergence was not better than ~ 10^-4 Hartree/cell even when 100 steps were used.
At the same time, the default settings of the INCAR at larger volumes (and, consequently, larger magnetic moments) provide a reasonable electronic convergence without flips of the local magnetic moments. In fact, the moments vs iteration curves are converged rather fast and, therefore, do not destabilize the minimization routine. Also no problems occur at small volumes where magnetic moments disappear.
I would like to ask whether someone else has experienced such nonconverging magnetic systems, and whether there exists a known solution of this issue. Any advise how to overcome the problem will be, of course, also highly appreciated. Thank you very much for your response.
With best regards,
Alexey Dick
P.S. Parallel/serial VASP v.4.6.31 compiled with Intel compiler was used.
The typical INCAR file that I have used is provided below:
PREC = accurate
ISMEAR = 1
SIGMA = 0.15
ENCUT = 270
ISPIN = 2
MAGMOM = 16*5.0 16*-5.0
ALGO = NORMAL
EDIFF = 1.0E-06
NELM = 100
NELMIN = 7
NELMDL = -7
ADDGRID = .TRUE.
ISTART = 0
ICHARG = 2
NBANDS = 200
NUPDOWN = 0
LMAXMIX = 6
AMIX = 0.2
AMIX_MAG = 0.1
BMIX = 0.0001
BMIX_MAG = 0.0001