A same deepks example can converge in the old version, but not in the new version

[xuan112358]

Describe the bug

I tried to use a deepks model to perform SCF calculation in a periodic system containing 32 water molecules in a unit cell.
When I used the older version of ABACUS, v3.1.2, it can converge normally.

But when I tried the newer version, 3.6.1, the DRHO would oscillate, and SCF could not converge.

I know the mixing method is different between these two versions. So I changed the method from "pulay" in v3.1.2 to "brodyen" in v3.6.2. Should I change other INPUT parameters to get converged results?
By the way, I have tested to use PBE in v3.6.2, and it can converge normally too.

Expected behavior

No response

To Reproduce

Inputs and results by v3.1.2 / v3.6.1 are in directory 01_3.1.2 / 02_3.6.2
github_files.zip
.

Environment

I use mpiicpc and icpc to compile.

Additional Context

No response

Task list for Issue attackers (only for developers)

• 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).

[WHUweiqingzhou]

@xuan112358,
I'd like to ask which version of ABACUS's data was used to train the DEEPKS model here? I suspect it's due to version 3.4.0, where we modified the two-center integrals @jinzx10, causing inconsistencies in the dataset precision and leading to a lack of convergence.
@dyzheng, what is your opinion?

[xuan112358]

@WHUweiqingzhou I used the data generated by ABACUS v3.1.2. Would this caused the problem?

[WHUweiqingzhou]

@xuan112358
Yes, I think that's one of the reasons. @dyzheng @wenfei-li @ouqi0711 what is your suggestion?

[ouqi0711]

I think the modification to the two-center integrals should be the reason that causing the convergence failure. We have investigated similar problem of the perovskite system. The short answer is that data/model trained via v3.4 or earlier can not be directly applied to the current version.

[xuan112358]

OK，I think I understand.
I also tried to compile ABACUS v3.6.2 without the new two-center integral code, by turning off the compilation option USE_NEW_TWO_CENTER. And I found that this time the result was consistent with the older version. I think it does suggest that it's the new two-center integrals that caused the different results.