Efficiency between VASP and QE for water

[zhaoru_sun1]

Dear developers,

I am using vasp for scf calculation, but I found a not small difference in efficiency between VASP and QuantumEspresso. I would like to know is there any way to improve the efficiency?
All the related files have been attached for reference.

• Test system: 64 H2O
• XC: SCAN
• CPU: 32
• Compiler: Intel oneapi 2021.5
• VASP: vasp.6.3.1 gamma-only
• QE: PWSCF v.7.0

Time cost

• VASP: 24 min
• QE: 6 min

[martin.schlipf]

The energy cutoffs are very large. I would recommend choosing a better pseudopotential in QE. It is probably possible with state of the are pseudopotentials to cut the energy cutoff in half. For VASP, 700eV would be recommended for a first run and you probably would need ~900eV for a production run.
How did you choose the energy cutoff?

[zhaoru_sun1]

The energy cutoffs are very large. I would recommend choosing a better pseudopotential in QE. It is probably possible with state of the are pseudopotentials to cut the energy cutoff in half. For VASP, 700eV would be recommended for a first run and you probably would need ~900eV for a production run.
How did you choose the energy cutoff?

Thank you. The cutoffs were chosen following some related references, which have calculated similar system and used the same XC. For example:
https://www.nature.com/articles/s41467-022-28538-8
https://journals.aps.org/prl/abstract/1 ... 126.236001

[martin.schlipf]

The energy cutoff is strongly dependent on the particular pseudopotential used. The norm-conserving ones you use for QE require significantly larger cutoffs than the PAW used by VASP. You should investigate a property of interest w.r.t. the energy cutoff. In VASP, start by setting ENCUT to the maximum ENMAX value in the POTCAR. Then increase the cutoff until the property you are interested in does not change anymore.

[zhaoru_sun1]

The energy cutoff is strongly dependent on the particular pseudopotential used. The norm-conserving ones you use for QE require significantly larger cutoffs than the PAW used by VASP. You should investigate a property of interest w.r.t. the energy cutoff. In VASP, start by setting ENCUT to the maximum ENMAX value in the POTCAR. Then increase the cutoff until the property you are interested in does not change anymore.

This is my test results for energy, the relative refer to the energy minus the largest cutoff result.
VASP

Code: Select all

Cutoff(eV)  Energy (eV)     Relative (eV)
700         -1031.94594795  1.93038155
800         -1033.2235293   0.6528002 
900         -1033.46395981  0.41236969
1000        -1033.53365312  0.34267638
1100        -1033.58233841  0.29399109
1200        -1033.61051068  0.26581882
1300        -1033.71857317  0.15775633
1400        -1033.78512778  0.09120172
1500        -1033.8763295   0. 

QE

Code: Select all

Cutoff(Ry)  Energy (Ry)     Relative (eV)
100         -2201.42976295  1.42929998 
110         -2201.49562590  0.53318897 
120         -2201.52348624  0.15412977 
130         -2201.53250960  0.03136071 
140         -2201.53563228  -0.01112551
150         -2201.53481457  0.

Also attach the test files.

[martin.schlipf]

Please do not converge the total energy. The total energy is not a meaningful number in pseudopotential DFT. Always look at the property you are actually interested in e.g. band gap to judge whether a system is converged. Most of this time the actual quantity of interest depends on energy differences which converge much more rapidly than the total energy.

[zhaoru_sun1]

I want to converge atomic forces. I set the average value of the force deviation of the O atoms as the convergence metric, because the force deviation of the H atoms is much smaller than that of the O.
When I use 2000eV as the standard, I see a jump of about 0.02eV/Angstrom between 1900eV and 2000eV. There is also about 0.02eV/Angstrom between 1600eV and 1500eV or 1700eV when I use 1600eV as the standard. So I think the final force deviation may be not from cutoff. But I'm not sure how to converge atomic force.

Use 2000eV as standard

Code: Select all

Cutoff(eV) AveForceDevi-O (eV/angstrom)
700 0.04980648070546598
800 0.03405969077756813
900 0.02834858094104418
1000 0.028791098611547022
1100 0.023417197174262036
1200 0.02621166934335298
1300 0.022773826881560322
1400 0.022508783417954564
1500 0.02269866950441226
1600 0.02117677697388945
1700 0.02005124092239465
1800 0.01968912679825246
1900 0.020036136422652215
2000 0.0

Use 1600eV as standard

Code: Select all

Cutoff(eV) AveForceDevi-O (eV/angstrom)
700 0.05004908294495167
800 0.035939357020531734
900 0.030175471467771893
1000 0.02986348230812608
1100 0.025017232272720624
1200 0.027212761268882887
1300 0.024693950575874035
1400 0.024052022916598872
1500 0.02469043757520198
1600 0.0
1700 0.02325924023174367
1800 0.021599347171176078
1900 0.02168993685773549
2000 0.02117677697388945

Calculation data and force deviation calculation code are attached.

[martin.schlipf]

What is you final goal, what do you need the forces for? Note that your EDIFF setting might not be sufficient for forces more accurate than what you observe.