ABINIT Discussion Forums

Hello:

I am trying to optimize the geometry of a crystal with 118 atoms in the unit cell with ionmov = 2. Unfortunately, this is exceedingly slow using the number of planewaves and k-points that seems necessary to get converged results, as described in the tutorials.

My question is, is it worthwhile (in terms of time and accuracy) to optimize the geometry with less planewaves or k-points first and then do a second optimization with the larger basis set using the geometry from the more coarse calculation?

Thank you,
Kale

Hello Kale

It depends on your system. I guess that would work for a calculation with standard DFT. You can start from a less converged calculation and increase the accuracy little by little. I would not expect this to work with a DFT+U calculation for instance, because the low converged calculation would get you in some metastable state in which your system will be stuck, even with higher convergence criteria.

For a large supercell, the best way would be to try and optimize npkpt, npband and bandpp. I guess you have very few kpoints and a lot of bands. Try npkpt = nkpt * nsppol, then adjust npband. If you increase bandpp, your scf steps will be slower but your convergence will be greatly improve, so in the end you get the same running time.

Good luck

Boris

I agree with Boris--kgb parallelization does work well for ground state calculations, including geometry optimization. My strategy is to take advantage of the paral_kbg -nproc feature, and see what abinit recommends for npkpt and so forth. You can then also tweak the k point grid (using prtkpt 1 to get different possibilities consistent with your system) and also nband, to force say a few extra empty bands so that the number of bands and k points both share a lot of common factors with the number of processors you can use, for best efficiency.

Dear Abinit Folks,

I too am trying to optimize the geometry of a crystal, however for a system with 109 atoms : a 3x3x3 fcc Al cluster with a self interstitial. I am doing this study for the system subjected to volumetric strain from 0% to 25% Tensile. The combination of ionmov=1, amu=28, dtion=400,vis=100 works robustly for low level of strain (0 - 12%). HOwever, beyond this level, I am unable to reduce the max force below 1e-3. I have tried out all the available ionmov schemes (BFGS, DIIS, Quenched MD, damped MD) but to no avail. I would be greatly obliged if you could provide me with some insight/ share your experience in relaxing such systems. I am also attaching my input file (if it is of any help)

Thanks,
Mrinal

Dear Folks,

I take back what I said. It turns out that the issues I was observing was because of not performing a k-point convergence study each time I increased the cell size. The errors in the force evaluation became too high (~1e-3) for any relaxation to occur.

Thanks,

Mrinal