Bug: cell-relax and SCF problem in symmetry=1 for ferromagnetic system

[QuantumMisaka]

Describe the bug

Example from @kirk0830 : CoFe2O4 cubic bulk, which is a ferromagnetic system.

ATOMIC_SPECIES
Fe   55.845  Fe_ONCV_PBE-1.0.upf
Co   58.933  Co_ONCV_PBE-1.0.upf
O   15.999  O_ONCV_PBE-1.0.upf

NUMERICAL_ORBITAL
Fe_gga_8au_100Ry_4s2p2d1f.orb
Co_gga_8au_100Ry_4s2p2d1f.orb
O_gga_7au_100Ry_2s2p1d.orb

LATTICE_CONSTANT
1.889726

LATTICE_VECTORS
8.563292000000  0.000000000000  0.000000000000
0.000000000000  8.563292000000  0.000000000000
0.000000000000  0.000000000000  8.563292000000

ATOMIC_POSITIONS
Direct

Fe
2.000
16
0.375000000000  0.875000000000  0.125000000000  1  1  1  
0.375000000000  0.125000000000  0.875000000000  1  1  1  
0.125000000000  0.875000000000  0.375000000000  1  1  1  
0.125000000000  0.125000000000  0.625000000000  1  1  1  
0.375000000000  0.375000000000  0.625000000000  1  1  1  
0.375000000000  0.625000000000  0.375000000000  1  1  1  
0.125000000000  0.375000000000  0.875000000000  1  1  1  
0.125000000000  0.625000000000  0.125000000000  1  1  1  
0.875000000000  0.875000000000  0.625000000000  1  1  1  
0.875000000000  0.125000000000  0.375000000000  1  1  1  
0.625000000000  0.875000000000  0.875000000000  1  1  1  
0.625000000000  0.125000000000  0.125000000000  1  1  1  
0.875000000000  0.375000000000  0.125000000000  1  1  1  
0.875000000000  0.625000000000  0.875000000000  1  1  1  
0.625000000000  0.375000000000  0.375000000000  1  1  1  
0.625000000000  0.625000000000  0.625000000000  1  1  1  

Co
-2.000
8
0.250000000000  0.750000000000  0.750000000000  1  1  1  
0.000000000000  0.000000000000  0.000000000000  1  1  1  
0.250000000000  0.250000000000  0.250000000000  1  1  1  
0.000000000000  0.500000000000  0.500000000000  1  1  1  
0.750000000000  0.750000000000  0.250000000000  1  1  1  
0.500000000000  0.000000000000  0.500000000000  1  1  1  
0.750000000000  0.250000000000  0.750000000000  1  1  1  
0.500000000000  0.500000000000  0.000000000000  1  1  1  

O 
0.000
32
0.135560000000  0.135560000000  0.864440000000  1  1  1  
0.135560000000  0.364440000000  0.635560000000  1  1  1  
0.364440000000  0.135560000000  0.635560000000  1  1  1  
0.364440000000  0.364440000000  0.864440000000  1  1  1  
0.114440000000  0.114440000000  0.385560000000  1  1  1  
0.114440000000  0.885560000000  0.614440000000  1  1  1  
0.385560000000  0.114440000000  0.114440000000  1  1  1  
0.385560000000  0.885560000000  0.885560000000  1  1  1  
0.135560000000  0.635560000000  0.364440000000  1  1  1  
0.135560000000  0.864440000000  0.135560000000  1  1  1  
0.364440000000  0.635560000000  0.135560000000  1  1  1  
0.364440000000  0.864440000000  0.364440000000  1  1  1  
0.114440000000  0.614440000000  0.885560000000  1  1  1  
0.114440000000  0.385560000000  0.114440000000  1  1  1  
0.385560000000  0.614440000000  0.614440000000  1  1  1  
0.385560000000  0.385560000000  0.385560000000  1  1  1  
0.635560000000  0.135560000000  0.364440000000  1  1  1  
0.635560000000  0.364440000000  0.135560000000  1  1  1  
0.864440000000  0.135560000000  0.135560000000  1  1  1  
0.864440000000  0.364440000000  0.364440000000  1  1  1  
0.614440000000  0.114440000000  0.885560000000  1  1  1  
0.614440000000  0.885560000000  0.114440000000  1  1  1  
0.885560000000  0.114440000000  0.614440000000  1  1  1  
0.885560000000  0.885560000000  0.385560000000  1  1  1  
0.635560000000  0.635560000000  0.864440000000  1  1  1  
0.635560000000  0.864440000000  0.635560000000  1  1  1  
0.864440000000  0.635560000000  0.635560000000  1  1  1  
0.864440000000  0.864440000000  0.864440000000  1  1  1  
0.614440000000  0.614440000000  0.385560000000  1  1  1  
0.614440000000  0.385560000000  0.614440000000  1  1  1  
0.885560000000  0.614440000000  0.114440000000  1  1  1  
0.885560000000  0.385560000000  0.885560000000  1  1  1  

When use symmetry 1 to do calculation, the SCF-step will be faster,
but it is hard for SCF to converge to drho=1e-6.
And when run cell-relax job, the BFGS optimization will break down.

While use symmetry 0, SCF will be slower, but can have a much much better convergence.

Expected behavior

symmetry 1 should have likely cell-relax and SCF convergence of symmetry 0, and also do quicker SCF calculation.

To Reproduce

• build ABACUS with intel-oneAPI toolchain
• run by OMP_NUM_THREADS=4 mpirun -np 16 abacus

Environment

• Intel-8358
• Rockylinux 8.5
• ABACUS 3.4.0 (3.4.1 can reproduce)
• Intel-MKL, compiled by icpx

Additional Context

input and result

CoFe2O4.tar.gz

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).

[QuantumMisaka]

in ABACUS 3.4.1 (Commit: 77f178d (Thu Oct 26 11:32:59 2023 +0800)), the problem above can be reproduced

3.4.1-result.tar.gz

[maki49]

I've reproduced the problem. It's not the fault of module symmetry itself.

The problem lies in poor symmetry keeping during the relaxation: the symmetry is broken after the 1st ion step from T_d to C_2v point group (I found it by running scf with STRU_ION2_D), but in the 2nd ion step ABACUS doesn't reanalyze the symmetry and still uses T_d to symmetrize the charge density, which possibly leads to non-convergence.

This will be fixed in the future by implementing #2901 after a refactor of UnitCell.

[hongriTianqi]

Would you refactor the module_cell? @maki49

[maki49]

Would you refactor the module_cell? @maki49

Yes. I want to make Symmetry a submodule of UnitCell to call it in module_relax. But now Symmetry itself depends on the entire class UnitCell. So I want to divide UnitCell into some submodules, firstly "lattice", "atoms" and "magnetic" on which Symmetry depends (too many parameters to pass if not packed as submodules).

[maki49]

It seems #3247 can fix this issue. Here's the log of my local test:
CoFe2O4_3247.log