My Community

Dear all,
I'm doing soc self-consistent calculation with afm ordering but it does not converge at all. I've tried all methods online I could find but none of them works. EDIFF should be 10E-8. But to get convergence first, I set it to be 10E-6 for now. Attached are files following the procedure in the attachment. First do ICHARG=12 and then do ICHARG=2 with mixing parameter.

For other methods:
1. direct computing soc using ICHARG=2 and tuning mixing parameters (amix from 0.2 to 0.01);
2. run non magnetic or collinear case first, and then set ICHARG=1 to read CHARGCAR of previous step to do soc calculation;
3. follow steps on vaspwiki https://www.vasp.at/wiki/index.php/Trou ... onvergence;
4. run gamma point only first and then read CHGCAR to run with normal k points. (all methods include tuning mixing parameter and setting PREC=Accurate, ADDGRID = .TRUE.).

None of these works.

Thank you for your help.

Dear yujia_teng,

If you have problems with converging your ab initio calculations it is advisable to try different settings for ALOG.
Right now I would try to redo your calculations with ALGO=All. Maybe this resolves your issue and otherwise, you can try other settings for the ALGO tag.
I will move the post to the using-vasp forum since this is not necessarily a bug yet.

If you still experience problems please contact us again.

All the best Jonathan

Dear admin,
I've done lots of tests with different ALGO, but it still does not converge well. It does not converge at all in most of my tests. In a few tests, where I set a NELM=100 or 200, convergence can be achieved with near the limit of electronic steps. But I don't think this is a good convergence or even a convergence, because among those a few converged tests, there're too many electronic steps there and usually convergence should be achieved within default electronic steps. Also, the resulted energies are different by order of 1～10 meV.

Dear yujia_teng,

As long as your calculation converges the results you obtain are fine. It does not matter how many steps you need until you reach convergence. But if you intend to improve the convergence of your calculations you could try to adjust the values of the density mixing parameters AMIX, BMIX and the mixing method. Another parameter that you could try to adjust is LMAXMIX.
When checking your INCAR file I recognized you are using a rather large value for BMIX. Maybe you can try the mixing parameters suggested in this post.
The attached slides might also be helpful for you. In these, the density mixing in vasp is explained.

All the best Jonathan

Dear admin,
It's good to know that result is fine as long as it gets convergence. The large BMIX=3.0 is from page 22 in the slides, the one you attached.

The mixing parameter in the forum post with that AMIX and BMIX value could give me convergence when EDIFF = 1e-06. But it still cannot converge if I set EDIFF = 1e-08（same for all other methods I can find online）. dE fluctuates between order -03 and -04 in this case. I'm not confident that I can get convergence with this higher accuracy. So in this kind of hard to converge case, the accuracy is always limited? Although 1e-06 usually is enough, in some magnetic case 1e-08 may be needed and some paper uses this value.

Dear yujia_teng,

You are right, some systems will be harder to converge than others. As in your case you recognize this by not being able to reach a convergence criterion of 1E-8. But still, you can try to optimize your electronic minimization by scanning your input parameters. For example, you could try to find optimal values for the density mixing for the considered system. You could also try to increase the ENCUT which might give you better electronic convergence.
I was checking your KPOINTS file again. I recognized you are using a 12x12x6 kpoints grid. The ratio between your lattice parameter c/a is roughly 1.3. So I would probably use more kpoints in the c direction. Maybe you could try a 12x12x12 grid to check if this improves your electronic convergence.

All the best Jonathan

Dear admin,
Thanks for your clarification. Writing a script scanning lots of input parameters is a better choice than try some manually. I would try this in the future.

I tried kpoints with 12*12*9, the convergence is indeed improved, but the energy is different from original 12*12*6 one by 5meV. (12*12*12 does not give convergence) So I still have one concern here: Is this difference of energy among different converged results acceptable?

Dear yujia_teng,

With a different number of kpoints it is expected to get a slightly different result for the electronic energy. The magnitude of the difference will depend on the system and the difference in the number of kpoints. In principle, one should always check a result with respect to the number of k-points. It is advisable to perform several calculations, gradually increasing the number of k-points, to ensure convergence. What error energy error is acceptable will depend on the properties that you want to compute and what accuracy is needed for these properties.

All the best Jonathan