Dear Abinit users,
I am dealing with a doped supercell system (Sr[Li2Al2O2N2] where one Sr atom is replaced by Eu atom). I treat it using GGA-PBE and PAW+U on 4f states of Eu. Those 4f states are localized and located in the band gap.
From previous computations with rather small supercell sizes, it appeared that the property that I need is hard to converge with respect to the distance between Eu images along c-axis. This is why I am looking for increasing the supercell size only along this c-axis.
With “small” supercell, from 54 (1x1x3) to 144 atoms (1x1x8), SCF cycles converge. The number of steps required to obtain SCF convergence increases with the supercell size (and therefore the Eu-Eu distance). It takes around 50 steps for 54 atoms (without help) and 100-150 steps for 144 atoms (with the help of nline=6, nnsclo =2, tolrde=0.001).
But here is my problem: with a 216 atoms supercell (1x1x12), I never succeeded to obtain convergence. I attach a typical input file and the related output.
Here are the input parameters I already tried to tune without success:
- increasing the k-points density (from one to 8 k-points)
- increasing ecut up to 45 Ha
- decreasing diemix (0.8, 0.6, 0.4)
- increasing nline (up to 15) and increasing nnsclo (up to 4)
In each case, the SCF cycle starts to converge during a few dozen of steps (the convergence of these steps improves especially when increasing nline) but then it starts diverging and it never stabilizes as can be seen in the output attached.
iter Etot(hartree) deltaE(h) residm nres2 diffor maxfor ETOT 1 -2559.1858774541 -2.559E+03 5.407E-07 4.153E+03 3.789E+00 3.789E+00 ETOT 2 -2665.5907405032 -1.064E+02 2.347E-07 2.554E+04 4.784E+00 9.944E-01 ETOT 3 -5172.4502902935 -2.507E+03 1.813E-03 8.728E+04 2.950E+00 3.043E+00 ETOT 4 -2954.4527677927 2.218E+03 5.870E-07 3.285E+04 3.904E+00 1.454E+00 ETOT 5 -2713.2064613295 2.412E+02 4.890E-07 1.879E+04 1.353E+00 7.001E-01 ETOT 6 -2594.6850567304 1.185E+02 3.655E-08 8.048E+03 1.014E+00 3.140E-01 ETOT 7 -2584.4804851793 1.020E+01 1.824E-06 8.292E+03 3.326E-01 2.258E-01 ETOT 8 -2571.4035516666 1.308E+01 6.945E-09 5.488E+03 3.241E-01 1.929E-01 ETOT 9 -2543.4688664001 2.793E+01 2.076E-09 9.583E+02 2.431E-01 7.863E-02 ETOT 10 -2751.9181227293 -2.084E+02 5.180E-05 2.957E+04 6.568E-01 6.290E-01 ETOT 11 -2829.0589438146 -7.714E+01 5.299E-08 3.635E+04 1.766E+00 1.363E+00 ETOT 12 -2572.1695476902 2.569E+02 2.246E-07 8.819E+03 1.597E+00 4.111E-01 ETOT 13 -2609.5407521394 -3.737E+01 7.304E-07 1.684E+04 3.686E-01 5.396E-01 ETOT 14 -2632.0929092042 -2.255E+01 3.803E-07 1.862E+04 4.049E-01 5.741E-01 ETOT 15 -2654.7546906833 -2.266E+01 9.001E-09 1.880E+04 3.116E-01 4.929E-01 ETOT 16 -2583.1252916852 7.163E+01 2.275E-08 1.122E+04 3.102E-01 5.249E-01 ETOT 17 -2842.7548183094 -2.596E+02 1.209E-07 3.462E+04 7.128E-01 7.216E-01 ETOT 18 -3358.1430785262 -5.154E+02 1.798E-04 5.078E+04 7.112E+05 7.112E+05 ETOT 19 -4054.4991717184 -6.964E+02 5.994E-05 9.163E+04 7.116E+05 2.718E+04 ETOT 20 -2649.9525640974 1.405E+03 3.447E-06 1.983E+04 2.718E+04 9.666E-01 ETOT 21 -4027.0801399053 -1.377E+03 3.385E-05 8.646E+04 2.483E+00 2.268E+00 ETOT 22 -3430.8508720718 5.962E+02 2.077E-04 6.033E+04 1.916E+00 2.102E+00 ETOT 23 -3751.0958226289 -3.202E+02 4.037E-05 6.176E+04 2.700E+00 2.592E+00 ETOT 24 -3681.5626676015 6.953E+01 8.603E-06 6.092E+04 1.937E+00 2.586E+00 ETOT 25 -5199.9703999524 -1.518E+03 2.172E-04 8.447E+04 5.535E+00 4.871E+00 ETOT 26 -3381.9883258386 1.818E+03 5.117E-06 6.323E+04 8.375E+05 8.375E+05 ETOT 27 -10978.488188589 -7.596E+03 1.629E-02 1.610E+05 7.470E+05 2.443E+05 ETOT 28 -4869.6765511203 6.109E+03 2.161E-03 8.626E+04 9.173E+05 9.181E+05 ETOT 29 -9126.6741376169 -4.257E+03 1.247E-02 1.511E+05 9.116E+05 7.047E+05 ETOT 30 -7893.9581529911 1.233E+03 8.796E-03 1.384E+05 7.047E+05 8.549E+00
Notice that the system is not yet relaxed. My initial structure guess comes from a relaxed 1x1x6 that I simply double along c-axis. I look for a DEN file that I will use as an input density for future structural relaxation.
Notice also that an undoped supercell with the same size converges nicely in a few dozen steps. It really seems that the issue comes from Eu states and that the distance between Eu images is important.
My questions are :
- Where does this convergence issue come from? I guess it is related to the 4f localized states located in the band gap … Can someone give hints about that?
- Indeed, do you have any ideas of what I could try to obtain convergence?