Bug: Geometry Optimization Problem in Li1Si64 example

[QuantumMisaka]

Describe the bug

In ABACUS Catalysis Practice Notebook:
https://nb.bohrium.dp.tech/detail/6316830243
The NEB example need to first do relax calculation for initial and final state

But for intial state in this example , when use CG method ,the force will be in flux:

 LARGEST GRAD (eV/A)  : 5.285e-02
 LARGEST GRAD (eV/A)  : 1.397e-01
 LARGEST GRAD (eV/A)  : 1.386e-01
 LARGEST GRAD (eV/A)  : 1.194e-01
 LARGEST GRAD (eV/A)  : 8.415e-02
 LARGEST GRAD (eV/A)  : 8.968e-02
 LARGEST GRAD (eV/A)  : 9.033e-02
 LARGEST GRAD (eV/A)  : 8.701e-02
 LARGEST GRAD (eV/A)  : 1.890e-01
 LARGEST GRAD (eV/A)  : 1.575e-01
 LARGEST GRAD (eV/A)  : 2.278e-01
 LARGEST GRAD (eV/A)  : 2.208e-01
 LARGEST GRAD (eV/A)  : 6.767e-01
 LARGEST GRAD (eV/A)  : 3.907e-01
 LARGEST GRAD (eV/A)  : 1.291e+00
 LARGEST GRAD (eV/A)  : 1.225e+00
 LARGEST GRAD (eV/A)  : 1.027e+00
 LARGEST GRAD (eV/A)  : 3.355e-01
 LARGEST GRAD (eV/A)  : 1.739e+00
 LARGEST GRAD (eV/A)  : 1.769e+00
 LARGEST GRAD (eV/A)  : 1.859e+00
 LARGEST GRAD (eV/A)  : 1.244e+00
 LARGEST GRAD (eV/A)  : 6.014e-01
 LARGEST GRAD (eV/A)  : 4.632e-01
 LARGEST GRAD (eV/A)  : 6.680e-01
 LARGEST GRAD (eV/A)  : 5.251e-01
 LARGEST GRAD (eV/A)  : 1.132e+00
 LARGEST GRAD (eV/A)  : 6.121e-01
 LARGEST GRAD (eV/A)  : 2.150e+00
 LARGEST GRAD (eV/A)  : 7.361e-01

And when use BFGS, the force will also fluctuate

 LARGEST GRAD (eV/A)  : 5.285e-02
 LARGEST GRAD (eV/A)  : 1.397e-01
 LARGEST GRAD (eV/A)  : 1.386e-01
 LARGEST GRAD (eV/A)  : 1.194e-01
 LARGEST GRAD (eV/A)  : 8.415e-02
 LARGEST GRAD (eV/A)  : 8.968e-02
 LARGEST GRAD (eV/A)  : 9.033e-02
 LARGEST GRAD (eV/A)  : 8.701e-02
 LARGEST GRAD (eV/A)  : 1.890e-01
 LARGEST GRAD (eV/A)  : 1.575e-01
 LARGEST GRAD (eV/A)  : 2.278e-01
 LARGEST GRAD (eV/A)  : 2.208e-01
 LARGEST GRAD (eV/A)  : 6.767e-01
 LARGEST GRAD (eV/A)  : 3.907e-01
 LARGEST GRAD (eV/A)  : 1.291e+00
 LARGEST GRAD (eV/A)  : 1.225e+00
 LARGEST GRAD (eV/A)  : 1.027e+00
 LARGEST GRAD (eV/A)  : 3.355e-01
 LARGEST GRAD (eV/A)  : 1.739e+00
 LARGEST GRAD (eV/A)  : 1.769e+00
 LARGEST GRAD (eV/A)  : 1.859e+00
 LARGEST GRAD (eV/A)  : 1.244e+00
 LARGEST GRAD (eV/A)  : 6.014e-01
 LARGEST GRAD (eV/A)  : 4.632e-01
 LARGEST GRAD (eV/A)  : 6.680e-01
 LARGEST GRAD (eV/A)  : 5.251e-01
 LARGEST GRAD (eV/A)  : 1.132e+00
 LARGEST GRAD (eV/A)  : 6.121e-01
 LARGEST GRAD (eV/A)  : 2.150e+00

and then break down:

 trust radius is too small! Break down.
 CHECK IN FILE : OUT.ABACUS/warning.log

But when use ASE-ABACUS to do optimization, force will normally converge

      Step     Time          Energy          fmax
BFGS:    0 13:51:20    -7073.123389         0.063117
BFGS:    1 13:52:37    -7073.123735         0.051192
BFGS:    2 13:53:49    -7073.124284         0.031589

The related files:
initial_stru_opt.tar.gz

likely problem
#2974

Expected behavior

The geometry relax job should be normally done by ABACUS itself within 3 steps like ASE-ABACUS do

To Reproduce

Download example, run it by ABACUS

Environment

• ABACUS version: 3.3.4
• Dependencies: Intel-OneAPI, ELPA, libxc
• System: AMD-3950x, Ubuntu 20.04 - WSL2
• Running: OMP_NUM_THREADS=16 mpirun -np 1 abacus

Additional Context

for ase-abacus, the optimization code is part of NEB code, so copy that and past here:

import os 
from ase.calculators.abacus import Abacus, AbacusProfile
from ase.optimize import FIRE, BFGS, QuasiNewton
from ase.io import read, write
#from pathlib import Path

# set pythonpath: not useful
# ROOTPATH=os.path.abspath("../..")
# os.environ['PYTHONPATH'] = f'{ROOTPATH}'

from abacus_neb import AbacusNEB

# setting
# optimizer = FIRE # suited for CI-NEB
optimizer = BFGS
init_directory = "INIT"
final_directory = "FINAL"
neb_directory = "OUT"
algorism = "improvedtangent" # IT-NEB is recommended
#dyneb=True  # default
interpolate = "linear" # linear or idpp
climb = True
n_max = 7
mpi = 1
omp = 16
abacus = 'abacus'
# example_dir = "/lustre/home/2201110432/example/abacus"
# pseudo_dir = f"{example_dir}/PP"
# basis_dir = f"{example_dir}/ORB"
pseudo_dir = "./src"
basis_dir = "./src"
pp = {"Li": "Li_ONCV_PBE-1.2.upf",
        "Si": "Si_ONCV_PBE-1.2.upf", }
basis = {"Li": "Li_gga_8au_100Ry_4s1p.orb",
            "Si": "Si_gga_8au_100Ry_2s2p1d.orb"}
kpts = [2, 2, 2]
parameters = {
    'calculation': 'scf',
    'xc': 'pbe',
    'ecutwfc': 100,
    'smearing_method': 'gaussian',
    'smearing_sigma': 0.002,
    'basis_type': 'lcao',
    'ks_solver': 'genelpa',
    'mixing_type': 'pulay',
    'scf_thr': 1e-6,
    'scf_nmax': 300,
    'out_chg': 1,
    'out_bandgap': 1,
    'kpts': kpts,
    'pp': pp,
    'basis': basis,
    'pseudo_dir': pseudo_dir,
    'basis_dir': basis_dir,
    'vdw_method': 'd3_bj',
    'cal_force': 1,
    'cal_stress': 1,
    'out_stru': 1,
    'out_chg': 0,
    'out_bandgap': 0,
}

os.environ['OMP_NUM_THREADS'] = f'{omp}'
profile = AbacusProfile(
    argv=['mpirun', '-np', f'{mpi}', abacus])

# Initial stru read from ABACUS, should do single point calculation
initial = read('./initial_stru', format='abacus')

# relax calculation by abacus
initial.calc = Abacus(profile=profile, directory=init_directory,
                     **parameters)
qn_init = optimizer(initial, trajectory='init_opt.traj')
qn_init.run(fmax=0.05)

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).

[QuantumMisaka]

Tried PW method in the advice of @kirk0830 ，BFGS will converge in 5 steps, CG will converge in 2 steps

initial_stru_opt_pw.tar.gz

It seems to be a LCAO problem in relax ?

Also, if turn of vdw_method d3, the force will directly below 0.05 eV/A and converge

[kirk0830]

#2974

[hongriTianqi]

@QuantumMisaka Do you have more updates on this issue?

[QuantumMisaka]

@QuantumMisaka Do you have more updates on this issue?

no more now

[pxlxingliang]

Hi, @QuantumMisaka, the bug in BFGS have been fixed in #3658
I have used the latest codes to run this example, and it can be converged in 9 ION steps.

https://app.bohrium.dp.tech/abacustest/?request=GET%3A%2Fapplications%2Fabacustest%2Fjobs%2Fjob-abacustest-v0.3.106-02c2bf