Magnetic Calculation(AFM)

[9yulo]

Dear TRIQS community,
I want to do an AFM calculation similar to the example you did on NdNiO2.

1-) Doing the NdNiO2 example this took me more than 24h running on 96 nodes. at the level of the analysis of the spectral function, to have the same tracing as here I have to do the maxent analysis?NiO or simply how to have this quantity?

2-) can we do the same AFM calculation with DFTTOOLS? because I launched the calculation in super slow I launched the calculation with solid_dmft (mpirun -np 240 solid_dmft confi.ini ) it has been running for more than 30 hours with a single iteration.

[the-hampel]

Hi @9yulo,

I will start answering your second question first. DFTTools is just a library with functions and classes. You have to write your own DMFT loop etc to actually use it. solid_dmft is exactly doing that, using DFTTools to write a DMFT loop. There should be very little overhead in your calculation caused by solid_dmft. Most of the time the code is spending in the solver call (cthyb). In the example you are mentioning roughly 85% of the time are spent there, making this the computational expensive part. However, there is nothing in solid_dmft slowing down your calculations compared to a more minimal setup you might achieve by tailoring the script to your problem. This should clarify a bit your second question.

Regarding your first question: the NdNiO2 tutorial example you are mentioning should take roughly 4-5 hours on 1024 cores. I am not sure how many cores you in the end used? You are mentioning 96 nodes? How many cores does each node have?

How computationally demanding your task is depends a lot on the system at hand. In general a larger U value will make things easier for cthyb and you can get better results with smaller number of ncycles. Or you can maybe reduce the number of orbitals in the active space, or use a less complicated local Hamiltonian structure. The example in the tutorial is rather "hard". For certain edge cases approximate solvers might also help. See for example tutorial 04-HubbardI-comparison.ipynb on https://github.com/TRIQS/tutorials/tree/3.1.x/AbinitioDMFT . However, solving the impurity problem within the DMFT loop numerically exact, i.e. with cthyb, is a much more demanding that typical DFT calculations. This should be the same for other DMFT implementations out there.

If you are interested in approximate solvers, we will integrate the latest https://github.com/TRIQS/hartree_fock/ solver in solid_dmft, which will be very fast solving DMFT on the Hartree level, which is the same approximation done in DFT+U.

Best,
Alex

[9yulo]

Hi @the-hampel,
Thank you for your enlightenment,

Yes indeed it was 96 cores and not nodes. but the calculation is well finished. and I still have some questions:

1-) for the parameter dc_type = 0 dc = True dc_dmft = False which is Type of double counting correction considered in DFTTOOLS we give the value of the occupation with DC_value. but with this method of solid_DMFT how to be sure that the electrons have converged? I have by this at the end of the calculation for the example of NdNiO2
total occupany of impurity 0: 8.5264
total occupany of impurity 1: 8.5264
what shows well an occupation of 8.5 for each impurity at the end of the DMFT calculation what is appreciably equal to the value obtained in the DFT calculation (OUTCAR)
total chargeù

Since in the example we are talking about the AFM calculation, I expected that at the end of the DMFT calculation the two SPINs would be opposite:

how to have the spectral function like here? NiO

[the-hampel]

The convergence can be best tested when looking at the observables_imp0.dat and conv_imp0.dat files. There you should check if convergence is reached over iterations.

The impurity occupations itself in DMFT will greatly depend on the chosen projection method. The orbital occupations in Vasp might be misleading. Those are straight projections for all KS states on atomic d orbitals in the PAW spheres. When using the dfttools vasp converter it will take the same projections, but depending on your chosen energy window only in specific window. This will loose some projection weight, i.e. occupation. More importantly the converter will orthonormalize the projectors, which Vasp is not doing. This will increase the number of electrons again. For the NdNiO2 example you should see around 8.0 occupation per Ni side in the d orbitals, whereas in solid_dmft you have 8.5 electrons. This is reasonably close.

I urge you to maybe also do the 4 tutorials here: https://github.com/TRIQS/tutorials/tree/unstable/AbinitioDMFT , which use solid_dmft in a very simple example, which you can run even on a normal laptop. It will also show you how to use MaxEnt to calculate spectral functions. For k-dependent spectral functions you have to use at the moment the interface through wannier90. This has technical reasons. However, the calculation of the impurity spectral function is explained in the tutorial I referred to.

Regarding the magnetic moment: solid_dmft only prints the occupation per spin channel. You have to calculate the magnetic moment as difference between n_up - n_down by yourself. In your case:

imp 0 spin up: 4.7519
imp 0 spin down: 3.7745

this would lead to a moment of ~1 mu_B. And you can see that with this definition imp0 has +1 mu_B, whereas imp1 has a moment of -1 mu_B.

Best,
Alex

[the-hampel]

Hi, yes we did. Everything should work when you apply the patches we provide here: https://github.com/TRIQS/solid_dmft/tree/unstable/Docker . In the folder vasp_diffs we provide patch files. You can apply them with the lines shown in the openmpi_dockerfile:

solid_dmft/Docker/openmpi_dockerfile

Lines 122 to 126 in fd41678

	# patch for Vasp CSC
	#COPY vasp_diffs /vasp/vasp_diffs
	#RUN cd /vasp/src \
	#&& for name in electron.F fileio.F locproj.F mlwf.F .objects; \
	#do patch $name -p1 -i ../vasp_diffs/$name; done

Solid_dmft will also work without these patches in projector mode. The changes are for CSC calculations, the CSC wannier90 interface, and some performance optimization. The version of the unstable branch work with Vasp 6.3.2 and the version on the 3.1.x branch of solid_dmft work with Vasp 6.3.0.

Best,
Alex

[9yulo]

Hi,
the patch and the compilation is well finished with but the LOCPROJ file is different than what I calculated with vasp.5.4.4.

VASP.5.4.4: LOCPROJ

orbital     1     1     1      -52.1286225553        1.0000000000
     1        0.0000002651       -0.0000000221
     2        0.0007498565       -0.0000623912
     3        0.0000039955       -0.0000003324
     4        0.0000430995       -0.0000035861
     5        0.0000004803       -0.0000000400
     6        0.0000001217       -0.0000000101
     7       -0.0001027510        0.0000085493
     8       -0.0000016500        0.0000001373
     9       -0.0003238348        0.0000269444
    10        0.0000000776       -0.0000000065

VASP.6.3.0: LOCPROJ

orbital     1     1     1      -52.1286225554        1.0000000000
     1       -0.0000002660
     2       -0.0007524477
     3       -0.0000040093
     4       -0.0000432485
     5       -0.0000004820
     6       -0.0000001221
     7        0.0001031061
     8        0.0000016557
     9        0.0003249537
    10       -0.0000000778

plovasp plo.cfg

Warning: could not identify MPI environment!
Starting serial run at: 2022-08-24 11:19:50.679546
Traceback (most recent call last):
  File "/lib/python3.8/runpy.py", line 194, in _run_module_as_main
    return _run_code(code, main_globals, None,
  File "/lib/python3.8/runpy.py", line 87, in _run_code
    exec(code, run_globals)
  File "/RIQS/install/lib/python3.8/site-packages/triqs_dft_tools/converters/plovasp/converter.py", line 89, in <module>
    main()
  File "/RIQS/install/lib/python3.8/site-packages/triqs_dft_tools/converters/plovasp/converter.py", line 86, in main
    generate_and_output_as_text(filename, vasp_dir)
  File "/RIQS/install/lib/python3.8/site-packages/triqs_dft_tools/converters/plovasp/converter.py", line 57, in generate_and_output_as_text
    vasp_data = vaspio.VaspData(vasp_dir, efermi_required=efermi_required)
  File "/RIQS/install/lib/python3.8/site-packages/triqs_dft_tools/converters/plovasp/vaspio.py", line 78, in __init__
    self.plocar.from_file(vasp_dir)
  File "/RIQS/install/lib/python3.8/site-packages/triqs_dft_tools/converters/plovasp/vaspio.py", line 141, in from_file
    self.proj_params, self.plo = self.locproj_parser(locproj_filename=vasp_dir + "LOCPROJ")
  File "/RIQS/install/lib/python3.8/site-packages/triqs_dft_tools/converters/plovasp/vaspio.py", line 301, in locproj_parser
    ctmp = complex(float(sline[1]), float(sline[2]))
IndexError: list index out of range

[the-hampel]

Hi,
somehow your LOCPROJ file is missing the complex part (2nd column) and I do not know why. I am not sure how your input looks like, but I just recompiled Vasp 6.3.2 with the patch files from the unstable solid_dmft branch, and rerun this example for SrVO3: https://github.com/TRIQS/solid_dmft/tree/3.1.x/doc/examples/svo-one-shot/converter and I obtain the following correct LOCPROJ:

     1   343    20     5   5.2197220  # of spin, # of k-points, # of bands, # of proj, Efermi
   ISITE:     2    R=     -0.5000000    -0.5000000    -0.5000000  PAW projector    :    dxy   
   ISITE:     2    R=     -0.5000000    -0.5000000    -0.5000000  PAW projector    :    dyz   
   ISITE:     2    R=     -0.5000000    -0.5000000    -0.5000000  PAW projector    :    dz2   
   ISITE:     2    R=     -0.5000000    -0.5000000    -0.5000000  PAW projector    :    dxz   
   ISITE:     2    R=     -0.5000000    -0.5000000    -0.5000000  PAW projector    :   dx2-y2 

orbital     1     1     1      -29.1215272684        1.0000000000
     1        0.0000002807       -0.0000000874
     2       -0.0000003833        0.0000003910
     3       -0.0000008653       -0.0000009103
     4       -0.0000001389        0.0000000433
     5        0.0000002804        0.0000006869
...

which looks the same for Vasp 5.4.4. Can you please try this example. If you don't get this result we can dig deeper.

Best,
Alex

[9yulo]

Hi,

with vasp.6.3.0, its work only with std version not gamma version of vasp.

Best

[the-hampel]

Honestly, I never tested the projectors with the gamma point only version of Vasp. Please ask in the Vasp developer forum why this produces different output. I guess this is because one can remove the phase of the wave function if one only has one k-point? But the output should not change in this case, hence this is a bug on the Vasp site.

[9yulo]

Hi,

• There an option to RESTART the calculation with Solid DMFT? if the calculation has stopped let say at the 13th iteration out of 15 for lack of time it is possible to continue the calculation but from the previous calculation?

• After finishing the example of the tutorial 4 what options should be used to use this on-shot calculation to make a self-consistent calculation.

• How to determine these two parameters? fit_min_w ; fit_max_w ( the max and the min matsubara frequency).

• I am curious to make a comparison between the Anti-ferromagnetic and ferromagnetic system of the example 4 which parameters to use? magnetic = True ; magmom = +0.01, +0.01 afm_order = False ??

[the-hampel]

Hi,

• to restart a calculation, just re-submit the same job, with the job output folder present. If the name of the jobfolder already exists in the current dir, solid_dmft will try to continue the calculation and add another n_iter_dmft cycles to it.
• Not sure what you mean with the second point? you have to enable the csc=True flag. Try to follow this example for the additional flag: https://triqs.github.io/solid_dmft/tutorials/Ce2O3_csc_w90/tutorial.html
• you have to do the calculation first with some educated guess for the tail fitting parameters. Then plot the resulting self-energy and manually use the tail fitting routines in triqs (an example can be found here: https://github.com/the-hampel/pystack/blob/a0b03d9db1e945ee0f94f07130a597d43e4152d1/plot_triqs.py#L37) to test different parameters to give a smooth self-energy for the system at hand.
• Yes, just can either put afm_order=False. An advanced option is to use afm_order=True and put the same magmom on the impurities, i.e.

magnetic = True 
magmom = +0.01, +0.01 
afm_order = True

This will enforce a FM solution on the two impurities. But it is always advisable to cross-check with a afm_order=False calculation if this is a stable solution (no symmetries used)

Best,
Alex