Request: Better convergence of HSE in magnetic system #7081
Replies: 16 comments
-
From my practice now, HSE calculation CANNOT be restarted from wfc file or restart file, they can only restart PBE part. |
Beta Was this translation helpful? Give feedback.
-
|
@dyzheng @PeizeLin May this problem need together view and work ? |
Beta Was this translation helpful? Give feedback.
-
|
I complete a HSE SCF calculation by using 'scf_the 1e-6' and Much time costs in EXX, and also force and stress calculation. I consider that there should be much space left for performance update. |
Beta Was this translation helpful? Give feedback.
-
|
Here are some tests of convergence steps.
It seems that in this system, gg0 does not affect the convergence speed. |
Beta Was this translation helpful? Give feedback.
-
But, there are two other things which are more important
|
Beta Was this translation helpful? Give feedback.
-
|
@QuantumMisaka, |
Beta Was this translation helpful? Give feedback.
-
There is some update, I'll have more discussion later |
Beta Was this translation helpful? Give feedback.
-
|
I am testing HSE computation performance on these FeCx systems below Calculation setting: And:
The test result till now is :
But Fe3C systems, even have less number of atoms, is hard to converge
My next step will test HSE performance in
Now I'm wandering:
I may need more discussion and cooperate for HSE usage in these Fe-contained magnetic system @PeizeLin @WHUweiqingzhou @mohanchen |
Beta Was this translation helpful? Give feedback.
-
|
I noticed that the |
Beta Was this translation helpful? Give feedback.
-
|
After normalizing KPT to But the time cost for |
Beta Was this translation helpful? Give feedback.
-
|
The Force and Stress calculation in HSE can be separated, if I turn down the Stress calculation, the time cost will be much lower. Then the time cost for HSE SCF can be accepted |
Beta Was this translation helpful? Give feedback.
-
|
But ,I still wonder can we use wavefunction extrapolation in HSE AIMD/Opt calculation? for the separated loop algorism will always do PBE first |
Beta Was this translation helpful? Give feedback.
-
There are only |
Beta Was this translation helpful? Give feedback.
-
|
@PeizeLin I wonder:
|
Beta Was this translation helpful? Give feedback.
-
|
After But HSE convergence in magnetic system is still a problem I'll keep an eye on it. |
Beta Was this translation helpful? Give feedback.
-
|
I will transfer this to discussion. Hopefully we will come up with a better solution in near future. |
Beta Was this translation helpful? Give feedback.


Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
-
Details
I've tested HSE SCF in magnetic system, example is Fe-bcc conventional cell:
And KPT is
9 9 9Fe-HSE.tar.gz
Information:
ABACUS version: 3.4.4: Commit: 5f9d472 (Mon Dec 4 14:10:21 2023 +0800)
Dependence: Intel-OneAPI and Intel-toolchain
LibRI and LibComm: latest version before Nov 18
At first, my INPUT example is
it is very hard to converge to
scf_the 1e-7, even cannot reachscf_thr 1e-6within 5-days calculation inOMP_NUM_THREADS=16 mpirun -np 4 abacusin Intel-8358And after I saw #3103 , I add a parameter in my INPUT:
After that, convergence performance is better, in 2-days calculation of
OMP_NUM_THREADS=24 mpirun -np 2 abacusin Intel-8162, the SCF converge toscf_thr 1e-6, but notscf_thr 1e-7And memory consumption is 50G during calculation. Is this performance normal and proper for this system ? Can some improvements be done ?
Also. there exists some problem from user for using HSE :
Updateing EXX and rerun SCFnotice), which will give user a bad view that the calculation is stuck. Can more print-out information like consumed time in EXX process and some key processexx_separate_loopis 0 or 1, if I directly use wfc or restart file from half-calculated HSE process, will the initialization useless because of the first PBE process ?Task list for Issue attackers (only for developers)
Beta Was this translation helpful? Give feedback.
All reactions