Automatic adjustment of 'symmetry' from 1 to 0

[scott-5]

Background

Sometimes when doing cell-relax calculations, several situations will occur like that:

 Refine the lattice parameters in STRU or use a different`symmetry_prec`. 
 CHECK IN FILE : OUT.ABACUS/warning.log

or

ERROR: Symmetry cannot be kept due to the lost of accuracy with atom position during cell-relax.
Please set `symmetry` to 0 or -1 in INPUT file.  

So we need to adjust symmetry_prec or symmetry manually. Especially when I need to lots of DFT calculations, for example in CALYPSO, in order to avoid manual intervention, I have to set symmetry to 0. But it costs more time.
running_cell-relax.zip

Describe the solution you'd like

If ABACUS can add an option that automatically symmetry switches from 1 to 0 when symmetry analysis goes wrong, that will save a lot of resources. Further, it would be better if it implement that when the accuracy is restored, symmetry dynamically turns back to 1.

Additional Context

No response

[maki49]

Good idea, seems not hard to implement. I'll have a try.
Could you please upload the working directories producing these 2 error logs? I will use them as test cases for my implementation.

[scott-5]

Sure!
case.zip

[maki49]

@scott-5 There are 2 different kinds of errors:

1. Error-exit at a too-large symmetry_prec (case 2). I've implemented your idea in Feature: New method to dynamically change symmetry_prec during cell-relax #2736 , with a pdf file for details and test results.
2. Mismatch of bravis lattice (case 1). It only happens in 1st ion step and gives the same error message as VASP, so I didn't deal with it. Here's a related discussion: symmetry_prec problem in ABACUS #2592.
   • If automatically setting symmetry to 0 when mismatch is much more convenient in practice, I'd like to add a new bool parameter symmetry_autoclose (default: false) to control whether "error-exit" or "close-symmetry-and-continue" when mismatch, and update corresponding docs and tests in another PR. How do you think of this solution?

[scott-5]

@maki49 Thanks first for this pr, in which new algorithm is great.

About the solution for automatically setting symmetry to 0 when mismatch:
In fact, some initial structure is from the last relax-cell (STRU_ION_D) when gradient accuracy calculation. In original implementation, STRU_ION_D with low symmetry accuracy cause mismatch error in the next step beginning. With new symmetry_prec implementation, errors about mismatch may be avoided partly.
But mismatch may still occur when the last step ends with inaccuracy coincidentally. So I think the solution is feasible and robust(not automatically setting symmetry to 0 directly within the program, instead, leave an option symm_autoclose).

[maki49]

@scott-5 Thanks for your agreement. The next pr is coming soon.
BTW, I've patched up the missing results in the pdf file provided in #2736.

[scott-5]

Looking forward to the next pr. And there is a slight problem with figure in Example 2 in patched pdf file: green line seems started with symmetry_prec less than 10-3.

[maki49]

Looking forward to the next pr. And there is a slight problem with figure in Example 2 in patched pdf file: green line seems started with symmetry_prec less than 10-3.

It is because I didn't print the data in the beginning ion step. I'll fix the output in the next pr.
Here's the log after I fix the output, grep symmetry_prec in this log and then you can see it increases to 1.28e-3 in 2nd ion step.
running_cell-relax-example2-ion1-2.log