-
Notifications
You must be signed in to change notification settings - Fork 145
Description
Describe the bug
Example
A Au-Pd doped ZSM-5 model having H2O, OH and CH3 subgroup for adsorption, which have 299 atoms in a 20-20-13 Angstrom orthorhombic system.

Which is optimized by VASP before ( long ago, so in VASP calculation, this structure will also have 0.0557 max force in the first Ion step with IVDW=12 , which means vdw_method d3_bj in ABACUS ).
Performance
GOOD! ABACUS have much much more quicker SCF calculation efficiency from VASP. but there will come bad news.
NOTICE: Pd have 0.7 magmom, so this result is not accurate, the accurate result is added below
-------------------------------------------
STEP OF ION RELAXATION : 1
-------------------------------------------
START CHARGE : atomic
DONE(9.97348 SEC) : INIT SCF
ITER TMAG AMAG ETOT(eV) EDIFF(eV) DRHO TIME(s)
GE1 -1.30e-13 9.06e-13 -1.056626e+05 0.000000e+00 2.017e-01 3.160e+01
GE2 9.38e-13 2.01e-12 -1.059443e+05 -2.816879e+02 9.306e-02 3.094e+01
GE3 8.88e-12 1.01e-11 -1.059751e+05 -3.074302e+01 3.229e-02 3.066e+01
GE4 3.05e-10 3.07e-10 -1.059748e+05 2.167723e-01 6.967e-03 3.057e+01
GE5 3.30e-09 3.30e-09 -1.059754e+05 -5.912615e-01 4.298e-03 3.056e+01
GE6 4.67e-08 4.67e-08 -1.059755e+05 -7.448240e-02 2.895e-03 3.056e+01
GE7 1.04e-06 1.04e-06 -1.059755e+05 -1.648595e-02 1.492e-03 3.058e+01
GE8 1.97e-05 1.97e-05 -1.059755e+05 -1.082713e-02 9.213e-04 3.052e+01
GE9 3.79e-04 3.80e-04 -1.059755e+05 -8.332945e-03 5.675e-04 3.055e+01
GE10 8.35e-03 8.36e-03 -1.059755e+05 -3.938630e-03 3.894e-04 3.055e+01
GE11 1.30e-01 1.30e-01 -1.059755e+05 -1.597125e-03 3.870e-04 3.056e+01
GE12 3.82e-03 3.82e-03 -1.059756e+05 -3.690817e-03 2.713e-04 3.051e+01
GE13 1.44e-04 1.44e-04 -1.059756e+05 -1.740610e-03 1.989e-04 3.056e+01
GE14 -8.29e-05 8.29e-05 -1.059756e+05 -9.073505e-04 1.416e-04 3.054e+01
GE15 -1.29e-04 1.29e-04 -1.059756e+05 -6.161120e-04 9.477e-05 3.050e+01
GE16 -1.59e-05 1.59e-05 -1.059756e+05 -3.088634e-04 5.894e-05 3.054e+01
GE17 3.79e-05 3.80e-05 -1.059756e+05 -2.916646e-04 4.064e-05 3.053e+01
GE18 1.59e-05 1.59e-05 -1.059756e+05 -8.356315e-05 1.986e-05 3.052e+01
GE19 9.97e-06 1.00e-05 -1.059756e+05 -1.067000e-04 1.322e-05 3.051e+01
GE20 2.81e-05 2.82e-05 -1.059756e+05 -5.153060e-05 8.471e-06 3.062e+01
GE21 7.76e-04 7.76e-04 -1.059756e+05 5.537133e-07 5.939e-06 3.057e+01
GE22 2.75e-04 2.75e-04 -1.059756e+05 -4.353574e-05 5.798e-06 3.054e+01
GE23 1.07e-05 1.07e-05 -1.059756e+05 -3.220730e-05 3.799e-06 3.058e+01
GE24 4.87e-06 4.88e-06 -1.059756e+05 -1.419179e-05 2.181e-06 3.050e+01
GE25 3.77e-07 3.92e-07 -1.059756e+05 -8.950339e-07 1.353e-06 3.063e+01
GE26 9.60e-07 9.70e-07 -1.059756e+05 1.249310e-07 8.801e-07 3.092e+01
And in VASP, the first Ion step will cost 15188.50 sec. for 35 SCF step.
In the relax step of BFGS, the largest gradient will stuck to 0.08 eV/A for a long time:
(base) [2201110432@wm2-login01 CH3-OH-H2O-AuPd-ZSM5]$ grep GRAD abacus.log
LARGEST GRAD (eV/A) : 3.348e-01
LARGEST GRAD (eV/A) : 1.070e+00
LARGEST GRAD (eV/A) : 2.099e+00
LARGEST GRAD (eV/A) : 7.064e-01
LARGEST GRAD (eV/A) : 3.474e-01
LARGEST GRAD (eV/A) : 4.958e-01
LARGEST GRAD (eV/A) : 3.211e-01
LARGEST GRAD (eV/A) : 4.325e-01
LARGEST GRAD (eV/A) : 3.193e-01
LARGEST GRAD (eV/A) : 3.721e-01
LARGEST GRAD (eV/A) : 2.491e-01
LARGEST GRAD (eV/A) : 4.400e-01
LARGEST GRAD (eV/A) : 3.101e+00
LARGEST GRAD (eV/A) : 4.384e-01
LARGEST GRAD (eV/A) : 2.396e-01
LARGEST GRAD (eV/A) : 1.772e-01
LARGEST GRAD (eV/A) : 3.450e-01
LARGEST GRAD (eV/A) : 4.744e-01
LARGEST GRAD (eV/A) : 1.960e-01
LARGEST GRAD (eV/A) : 4.383e-01
LARGEST GRAD (eV/A) : 1.962e-01
LARGEST GRAD (eV/A) : 1.322e-01
LARGEST GRAD (eV/A) : 1.518e-01
LARGEST GRAD (eV/A) : 2.251e-01
LARGEST GRAD (eV/A) : 1.355e-01
LARGEST GRAD (eV/A) : 8.976e-02
LARGEST GRAD (eV/A) : 1.867e-01
LARGEST GRAD (eV/A) : 8.353e-02
LARGEST GRAD (eV/A) : 8.558e-02
LARGEST GRAD (eV/A) : 7.731e-02
LARGEST GRAD (eV/A) : 8.476e-02
LARGEST GRAD (eV/A) : 7.717e-02
LARGEST GRAD (eV/A) : 9.897e-02
LARGEST GRAD (eV/A) : 1.063e-01
LARGEST GRAD (eV/A) : 8.415e-02
LARGEST GRAD (eV/A) : 2.645e+00
LARGEST GRAD (eV/A) : 1.505e+00
LARGEST GRAD (eV/A) : 8.437e-02
LARGEST GRAD (eV/A) : 8.477e-02
LARGEST GRAD (eV/A) : 8.416e-02
LARGEST GRAD (eV/A) : 8.479e-02
LARGEST GRAD (eV/A) : 8.377e-02
LARGEST GRAD (eV/A) : 8.448e-02
LARGEST GRAD (eV/A) : 8.359e-02
LARGEST GRAD (eV/A) : 8.430e-02
LARGEST GRAD (eV/A) : 8.358e-02
LARGEST GRAD (eV/A) : 8.422e-02
LARGEST GRAD (eV/A) : 8.351e-02
LARGEST GRAD (eV/A) : 8.403e-02
LARGEST GRAD (eV/A) : 8.347e-02
LARGEST GRAD (eV/A) : 8.390e-02
LARGEST GRAD (eV/A) : 8.337e-02
LARGEST GRAD (eV/A) : 8.380e-02
LARGEST GRAD (eV/A) : 8.336e-02
LARGEST GRAD (eV/A) : 8.368e-02
LARGEST GRAD (eV/A) : 8.330e-02
LARGEST GRAD (eV/A) : 8.357e-02
LARGEST GRAD (eV/A) : 8.326e-02
LARGEST GRAD (eV/A) : 8.349e-02
LARGEST GRAD (eV/A) : 8.315e-02
LARGEST GRAD (eV/A) : 8.334e-02
LARGEST GRAD (eV/A) : 8.307e-02
LARGEST GRAD (eV/A) : 8.322e-02
LARGEST GRAD (eV/A) : 8.295e-02
LARGEST GRAD (eV/A) : 8.310e-02
LARGEST GRAD (eV/A) : 8.284e-02
LARGEST GRAD (eV/A) : 8.296e-02
LARGEST GRAD (eV/A) : 8.264e-02
LARGEST GRAD (eV/A) : 8.284e-02
LARGEST GRAD (eV/A) : 8.255e-02
LARGEST GRAD (eV/A) : 8.259e-02
LARGEST GRAD (eV/A) : 8.239e-02
LARGEST GRAD (eV/A) : 8.242e-02
LARGEST GRAD (eV/A) : 8.216e-02
LARGEST GRAD (eV/A) : 8.219e-02
LARGEST GRAD (eV/A) : 8.193e-02
LARGEST GRAD (eV/A) : 8.196e-02
LARGEST GRAD (eV/A) : 8.161e-02
LARGEST GRAD (eV/A) : 8.169e-02
LARGEST GRAD (eV/A) : 8.125e-02
LARGEST GRAD (eV/A) : 8.140e-02
LARGEST GRAD (eV/A) : 8.095e-02
LARGEST GRAD (eV/A) : 8.084e-02
LARGEST GRAD (eV/A) : 8.065e-02
LARGEST GRAD (eV/A) : 8.049e-02
LARGEST GRAD (eV/A) : 7.998e-02
LARGEST GRAD (eV/A) : 7.998e-02
LARGEST GRAD (eV/A) : 7.943e-02
LARGEST GRAD (eV/A) : 7.936e-02
LARGEST GRAD (eV/A) : 7.882e-02
LARGEST GRAD (eV/A) : 7.856e-02
LARGEST GRAD (eV/A) : 7.813e-02
LARGEST GRAD (eV/A) : 7.787e-02
LARGEST GRAD (eV/A) : 7.706e-02
LARGEST GRAD (eV/A) : 7.696e-02
LARGEST GRAD (eV/A) : 7.595e-02
LARGEST GRAD (eV/A) : 7.567e-02
LARGEST GRAD (eV/A) : 7.498e-02
LARGEST GRAD (eV/A) : 7.420e-02
LARGEST GRAD (eV/A) : 7.359e-02
LARGEST GRAD (eV/A) : 7.277e-02
LARGEST GRAD (eV/A) : 7.199e-02
LARGEST GRAD (eV/A) : 7.087e-02
LARGEST GRAD (eV/A) : 6.990e-02
LARGEST GRAD (eV/A) : 6.892e-02
LARGEST GRAD (eV/A) : 6.763e-02
LARGEST GRAD (eV/A) : 6.622e-02
LARGEST GRAD (eV/A) : 6.492e-02
LARGEST GRAD (eV/A) : 6.385e-02
LARGEST GRAD (eV/A) : 6.106e-02
LARGEST GRAD (eV/A) : 6.000e-02
LARGEST GRAD (eV/A) : 5.643e-02
LARGEST GRAD (eV/A) : 3.736e-02
In CG, the largest gradient will jump like zig-zag
LARGEST GRAD (eV/A) : 3.348e-01
LARGEST GRAD (eV/A) : 6.110e-01
LARGEST GRAD (eV/A) : 1.350e+00
LARGEST GRAD (eV/A) : 5.196e-01
LARGEST GRAD (eV/A) : 6.134e-01
LARGEST GRAD (eV/A) : 3.508e-01
LARGEST GRAD (eV/A) : 2.633e-01
LARGEST GRAD (eV/A) : 1.795e-01
LARGEST GRAD (eV/A) : 3.730e-01
LARGEST GRAD (eV/A) : 2.007e-01
LARGEST GRAD (eV/A) : 2.895e-01
LARGEST GRAD (eV/A) : 2.321e-01
LARGEST GRAD (eV/A) : 2.972e-01
LARGEST GRAD (eV/A) : 2.196e-01
LARGEST GRAD (eV/A) : 2.959e-01
LARGEST GRAD (eV/A) : 1.859e-01
LARGEST GRAD (eV/A) : 2.837e-01
LARGEST GRAD (eV/A) : 2.048e-01
LARGEST GRAD (eV/A) : 2.540e-01
LARGEST GRAD (eV/A) : 1.838e-01
LARGEST GRAD (eV/A) : 1.648e-01
LARGEST GRAD (eV/A) : 1.086e-01
LARGEST GRAD (eV/A) : 1.655e-01
LARGEST GRAD (eV/A) : 9.874e-02
LARGEST GRAD (eV/A) : 1.113e-01
LARGEST GRAD (eV/A) : 1.212e-01
LARGEST GRAD (eV/A) : 1.754e-01
LARGEST GRAD (eV/A) : 9.142e-02
LARGEST GRAD (eV/A) : 1.666e-01
LARGEST GRAD (eV/A) : 1.210e-01
LARGEST GRAD (eV/A) : 1.553e-01
LARGEST GRAD (eV/A) : 1.628e-01
LARGEST GRAD (eV/A) : 2.205e-01
LARGEST GRAD (eV/A) : 1.313e-01
LARGEST GRAD (eV/A) : 1.895e-01
LARGEST GRAD (eV/A) : 7.684e-02
LARGEST GRAD (eV/A) : 1.925e-01
LARGEST GRAD (eV/A) : 5.456e-02
LARGEST GRAD (eV/A) : 1.570e-01
LARGEST GRAD (eV/A) : 6.026e-02
LARGEST GRAD (eV/A) : 1.321e-01
LARGEST GRAD (eV/A) : 6.390e-02
LARGEST GRAD (eV/A) : 1.176e-01
LARGEST GRAD (eV/A) : 5.683e-02
LARGEST GRAD (eV/A) : 1.417e-01
LARGEST GRAD (eV/A) : 4.094e-02
Two of which will converge. but the relax performance looks poor.
Expected behavior
The relax process need to be more efficient and the relax trajectory should be more normal.
To Reproduce
Run the example above by BFGS or CG (CG_BFGS and FIRE do not try) by using scf_thr 1e-6 and force_thr_ev 0.05 and LCAO-std basis set.
Environment
- ABACUS version: 3.4.1 Commit: 77f178d (Thu Oct 26 11:32:59 2023 +0800)
- Dependencies: Intel-oneAPI toolchain
Additional Context
The relax ( and cell-relax ) module in ABACUS should be a widely-used module, which need more attention to be paid to optimize.
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).