Geometry Optimization Convergence Issues and Final Summary

[ismiek]

Hi everyone,
In the very last geometry optimization, which is provided with vtight convergence creep by default, none of the 19 or (in the case of another molecule 129) were successfully optimized. Now I would either need a tool that still displays a final summary of the calculation or how to set the convergence criteria of the individual geometry optimizations of the algorithm (e.g. after MTD or MD or after GC). Thank you in advance for your help.

[ismiek]

P.S. I know that I can change the criteria of the last geometry optimization with the option --optlev. However, my question was how I can change the other geometry optimizations or how I can write this in an input.toml and not as a command line.

[pprcht]

You can add the max number of optimization cycles, as well as the energy and gradient convergence criteria (in Hartree) manually in the [calculation] block, like in this example input: https://crest-lab.github.io/crest-docs/page/documentation/inputfiles_examples.html#geometry-optimization-custom
However, this will be applied to all optimizations.

If you don't want to run through the whole sampling again, you can try to put the last ensemble you got before the failure through the --mdopt runtype. A toml file can be set up like this: https://crest-lab.github.io/crest-docs/page/documentation/inputfiles_examples.html#ensemble-optimization

All optimizations failing sounds a bit like a version issue though. Are you using an ifort build build via CMake by any chance?

[ismiek]

First of all, thank you for your answer.
Setting the convergence criteria sounds interesting, but isn't there a way to select which geometry optimization gets which criteria in the algorithm?
I would like the last sampling to be performed. I was thinking about adding extreme convergence criteria instead of vtight.
I don't have enough knowledge for your last question. Maybe it will help you that I used CREST version 3.0 and that it worked with the system of Si4Cl10 and Si5Cl12 and that this problem only occurred with n-Si8Cl18 and neo-Si8Cl18.

[pprcht]

Setting the convergence criteria sounds interesting, but isn't there a way to select which geometry optimization gets which criteria in the algorithm?
I would like the last sampling to be performed. I was thinking about adding extreme convergence criteria instead of vtight.

That's exactly the optlev option you already mentioned yourself. Even more customization for the multistep optimization is not possible at the moment. Also the difference between vtight and extreme is not that significant, at least not for the system sizes you mention. Consider the maxcycle setting, if the optimizations terminated early the flaw is not with the convergence criteria (and it will not affect the other multilevel convergence settings).

I don't have enough knowledge for your last question. Maybe it will help you that I used CREST version 3.0

Unfortunately not. But since you didn't seem to have compiled the program with CMake yourself, I don't think it will have been this. If you want to make sure, run crest --dry and look at the binary information at the end of its output. It will tell you.
You may also want to consider updating to 3.0.1, there have been some hotfixes after reports from other users.

[ismiek]

Unfortunately, the maxcycle option, which I drove to 300, only delivered 1 conformer in this calculation of the n-Si8, which is not the point of CREST. I used an older Crest_Version (2.11.1) with default settings for the same molecule and here I got 540 conformers. So something is wrong with the new versions with the last geometry optimization, because until then both outputs look the same. Also the keywords --optlev extreme was not accepted. Should I send you the two outputs, if so via which platform?

[pprcht]

the maxcycle option, which I drove to 300

If true, you actually decreased the maxcycles with this. At vtight and extreme levels the maxcycles are max{20*N_at, 200} by default (cf. Table 5 of doi.org/10.1063/5.0197592), i.e. 520 for Si8Cl18.

So something is wrong with the new versions with the last geometry optimization, because until then both outputs look the same.

I disagree with this assessment, there are many things going on that are not printed and it can very well be caused by the xTB/tblite backend. We've noticed the SCC sometimes behaves differently.

Also the keywords --optlev extreme was not accepted.

You can add it in the toml as optlev='extreme' within the [calculation] block

Should I send you the two outputs, if so via which platform?

Attach them here, I'll see if I notice anything else

[ismiek]

2_11_1_nsi8.txt
3_0_1_nsi8.txt

[ismiek]

Thanks for the clarification on the default settings for the maxcycle.I tried entering the keyword via the TOML file, but with the same result that the last geometry optimization runs with vtight convergence criteria.

[ismiek]

The problem was solved by turning up the maximum cycles. Sometimes several calculations have to be carried out to obtain a meaningful result.