URGENT: Inconsistent stress and force calculated by ABACUS and QE, resulting different even unreasonable structures

[kirk0830]

Describe the bug

Combining with previous issues #2974 #3002 , a high-throughput calculation reports maybe-unreliable result produced by ABACUS's cell relax function:

x: ratio between final and initial volume of cell
y: ratio between the maxial volume during relaxation and initial volume of cell
One of examples, the alloy Li-Rh, its cell-relax calculation is tested for illustration. Starting structure is shown in following:

For the first estimations of stress tensors and forces calculated by QE and ABACUS, they are:
QE:

ABACUS:

, seems ABACUS deviates too much.
Input scripts and outputs are provided below, for this is a fundamental function, I suggest solve it ASAP.

Expected behavior

At least keep consistencies on stress and forces calculation results on identical structure between ABACUS and QE.

To Reproduce

Keeping identical pseudpotential files used, disabled DFT-D3 correction, ABACUS input script:
abacustest-vcrelax-LiRh.zip
QE:
qetest-vcrelax-LiRh.zip

Environment

Any

Additional Context

No response

Task list for Issue attackers

• Verify the issue is not a duplicate.
• Describe the bug.
• Steps to reproduce.
• Expected behavior.
• Error message.
• Environment details.
• Additional context.
• Assign a priority level (low, medium, high, urgent).
• Assign the issue to a team member.
• Label the issue with relevant tags.
• Identify possible related issues.
• Create a unit test or automated test to reproduce the bug (if applicable).
• Fix the bug.
• Test the fix.
• Update documentation (if necessary).
• Close the issue and inform the reporter (if applicable).

[dyzheng]

When you mentioned that "stress and force inconsistent", it suggested the energy is consistent.
But the energy is not consistent between ABACUS and QE : "ABACUS: -718.943 Ry ; QE: -718.182 Ry".
Please check the reason of energy inconsistent first.

[dyzheng]

I tested the case with k-points 222 to accelerate the calculation.
Energy of ABACUS is -718.137 Ry with setting "symmetry 0" and -718.895 Ry with setting "symmetry 1".
Energy of QE is -718.137 Ry with/without setting "nosym = .FALSE.".

[dyzheng]

I tested the case with k-points 2_2_2 to accelerate the calculation. Energy of ABACUS is -718.137 Ry with setting "symmetry 0" and -718.895 Ry with setting "symmetry 1". Energy of QE is -718.137 Ry with/without setting "nosym = .FALSE.".

Result of ABACUS is correct before #1904 , and turn to error ever since.
@maki49 can you help for this Issue?

[maki49]

Result of ABACUS is correct before #1904 , and turn to error ever since. @maki49 can you help for this Issue?

I've reproduced your result. It's the same problem as #2608: If charge density symmetrization is done in real space as default, the results are wrong; if it's done in reciprocal space, the results are right.
In QE, only reciprocal space charge density symmetrization is supported.

Now I have no better idea to fix the real-space symmetrization, my current suggestion is to do the charge density symmetrization in reciprocal space in any case like what QE does. (#3081 )

(Notice: I have never changed the real-space symmetrization algorithm, the right result before #1904 is because the symmetry analysis is incomplete.)

[dyzheng]

Thank for your reply, will the symmetry algorithm in ABACUS be completely consistent with QE after #3081? @maki49

[maki49]

Thank for your reply, will the symmetry algorithm in ABACUS be completely consistent with QE after #3081? @maki49

Not really.
The "symmetry analysis" algorithm is consistent with VASP, sometimes finding more symmetry operations (leading to fewer IBZ-k-points) than QE.
The “charge density symmetrization” algorithm is consistent with QE then, while VASP supports real-space symmetrization for pure point groups.

[pxlxingliang]

I plane to build a test suite for symmetry and use the energy/force/stress of symmetry=0 as the reference. Will check the energy difference (ED), maximum absolute difference of force (MADF), and maximum absolute difference of stress (MADS). How should I set the criteria? @maki49 @dyzheng
How about:
ED: < 1e-8 eV
MADF: < 1e-6 eV/A (the valid digits output in the log are 8 digits)
MADS: < 1e-4 kbar (the valid digits output in the log are 6 digits)

[maki49]

I plane to build a test suite for symmetry and use the energy/force/stress of symmetry=0 as the reference. Will check the energy difference (ED), maximum absolute difference of force (MADF), and maximum absolute difference of stress (MADS). How should I set the criteria? @maki49 @dyzheng How about: ED: < 1e-8 eV MADF: < 1e-6 eV/A (the valid digits output in the log are 8 digits) MADS: < 1e-4 kbar (the valid digits output in the log are 6 digits)

It's normal that symmetry=0 and 1 give slightly different energy (sometimes1e-3 Ry) due to the charge density symmetrization, and it may depends on ecutwfc and scf_thr. Here are some results I've tested before: https://xmywuqhxb0.feishu.cn/docx/MAAadv569onfU0xtbc5cEFwjnRh?from=from_copylink

[kirk0830]

Status update: this issue now is assigned to @Cstandardlib now.