Extreme ensemble-trimming by QCG results in a single output structure

[Physopholy]

Hello, everyone.

I updated from << 3.0 CREST to 3.0.2 and have since been experiencing some strange behavior from QCG. In case it's relevant, xTB is version 6.7.1. The issue occurs during "Ensemble optimization with tight thresholds," after cluster growth and the metadynamics runs. I'll have a few thousand structures in the rotamers xyz file, but then only one remains after relaxation.

I've had this issue with both organometallic complexes of ~60 atoms and for single metal ions. Here's my command:

crest input.xyz --qcg solvent.xyz --gfn2 --alpb h2o --chrg 2 --uhf 0 --notopo --ensemble --nopreopt --nsolv 14 --len x2.0 --mrest 20 --nclus 30 --esort --ewin 30 -T 128

Where input.xyz has been relaxed using the same settings with xtb immediately before, and the same for solvent.xyz, which is a water molecule. As you can see, I'd like about 30 clusters which I intend to refine with CENSO after. I had improved the issue using --esort and --ewin 30 (i.e., it was present before they were added), so I thought the problem was related to excessively tight RMSD tolerances, but the issue has returned for a new set of calculations I'm doing. It is not always present, i.e. when input.xyz is a Pb atom it is a problem, but for K (with chrg 1) it is not.

Can someone summarize what I need to do to get the ~30 clusters I've requested? Maybe I should add some absurd energy window like 1000 kcal/mol instead of 30 kcal/mol? This would only impact the energy sorting, correct?

Another point: sometimes I get an error during the same step: forrt1: severe (40): recursive I/O operation, unit -129 file unknown

Also, should I use the --nci setting? I had mixed results for it with the solvated organometallic clusters in the past, which are obviously a mix of covalent and non-covalent interactions.

Thanks!

[cplett]

Hi,
Thank you for reporting. Did the problem also occur with a previous crest version (<3.0)? If so, it's possible that the originally tested legacy conformer-search algorithm works as intended (it can be used by adding the --legacy keyword.
Additionally, it would be helpful if you could provide me with the input files so that I can check on this.

[Physopholy]

Hello, thanks for the suggestion - I had overlooked that flag.

Yes, it seems to work as expected for CREST 2.12 with xTB 6.4.0. Thankfully, I have been able to keep using your very useful tool for my research by just reverting to those versions. Unfortunately, the legacy flag didn't fix the issue.

Here is an example of a run with the issue.

edit: I guess the issue is due to the water molecule dissociation during --grow. This doesn't happen on CREST 2.x, for some reason.

[cplett]

Thanks to your detailed report, I was able to identify the issue. The problem was related to the charge not being set correctly for the QCG cluster growth in the current crest version. This bug is fixed with #401, and the example should work properly again.

[Physopholy]

Awesome! This is great news for me, as almost everything I'm using this tool for is charged and needs to be explicitly solvated. Thanks!