Minimum Hopping does not work in xtb 6.7.0 and 6.7.1

[PJpUnkie]

Dear XTB Team,

First of all, thank you for this great software, which I can enjoy even as an experimental chemist.

I frequently use Minimum Hopping, as it is sometimes superior to CREST. However, in the latest version (6.7.0), it doesn't work properly.

Describe the bug
In XTB 6.7.0, Minimum Hopping does not hop. Either the geometry optimization or the MD part doesn't start, as the trajectory consists of multiple structures with the same (starting) geometry and increasing energy (see the attached PNG). The problem is the same for both GFN-FF and GFN2 levels. This bug was observed in the precompiled binary (XTB version 6.7.0 (08769fc) compiled by 'albert@albert-system' on 2024-03-04) using table PC with Linux WSL Ubuntu. In the previous version (XTB version 6.6.1 (8d0f1dd) compiled by 'stahn@M-Bot' on 2023-08-01), it worked just fine.

See the comparison between versions 6.6.1 and 6.7.0

To Reproduce
In xtb 6.7.0 run
xtb geometry.xyz --metaopt --input inputfile.inp

Input file:
$write
output file=out.out
$metadyn
save=100
kpush=0.05
alp=1.2
$end

Here are prints of the calculations in verbose mode for both XTB 6.6.1 (which works fine) and XTB 6.7.0 (problematic) for both GFN2 and GFN-FF levels of theory.
Metaopt_test_xtb_6.7.0_GFN-FF.txt
Metaopt_test_xtb_6.7.0_GFN2.txt
Metaopt_test_xtb_6.6.1_GFN2.txt
Metaopt_test_xtb_6.6.1_GFN-FF.txt

Expected behaviour
The output of the calculation should be a trajectory of multiple optimized structures (*.xtbmeta.log), as it was in 6.6.1.

Thank you very much

[PJpUnkie]

The problem also prevails in the new version (xtb version 6.7.1 (edcfbbe) compiled by 'albert@albert-system' on 2024-07-22), minimum hopping still gives a set of identical geometries with increasing energy instead of a set of optimized conformers.

[oboril]

I can confirm this bug - just spend a couple hours figuring out why --metaopt does nothing more than just --opt (xtb version 6.7.1 (edcfbbe)). WIth gfn2 it just silently fails after the first local optimization, with gfnff it raises a warning, either way it does not perform any minimum hopping.

In both cases, error is not raised and the first local minimum is returned.

[PJpUnkie]

I found a way to set up minimum hopping to work in xTB version 6.7.1. In versions earlier than 6.6.1, it was only necessary to define $metaopt in the input file. However, in the latest version, $opt must be included in the input file.

There are two options that worked for me:

Set maxcycle=20 – This allows the optimization to proceed with a limited number of geometry optimization steps using the default engine. However, if maxcycle is set to a value greater than 30, it stops working (i.e., no conformers are generated).

Specify the optimization engine as either lbfgs or inertial with the default maxcycle (maxcycle=0). Both options work fine. When using engine=lbfgs, the results are almost comparable to the default optimization engine (ANCopt) used in version 6.6.1.

Interestingly, when the maxcycle=20 is used, the log file sometimes contains structures where the bias was applied, in addition to the local minima. However, with the other than default engine (or default engine in version 6.6.1) and maxcycle=0, copies of the last local minimum are saved instead of biased structures (which can be easily filtered through cregen from CREST). Also with FIRE engine less local minima are found, as more steps of bias addition are used between minima.

Examples of working input file:

$write  
   output file=out.out  
$metadyn  
   save=100  
   kpush=0.05  
   alp=1.2  
$opt  
   maxcycle=20  
$end 

or

$write  
   output file=out.out  
$metadyn  
   save=100  
   kpush=0.05  
   alp=1.2  
$opt  
   engine=inertial/lbfgs   
$end 

Here are total input and output files of metaopt with Aminocaproic acid at GFNFF level and metadyn set as

$write
   output file=out.out
$metadyn
   save=300
   kpush=0.05
   alp=1.2

ACA.xyz

22
Energy: -19.8344221
C 1.17566 -0.02979 0.31977
C 0.43158 1.05210 -0.46728
C -0.93726 1.36148 0.14937
C -1.26050 2.84057 0.14315
O -0.43895 3.74121 0.16755
O -2.59174 3.08481 0.17149
C 2.56970 -0.29445 -0.25883
C 3.17567 -1.56776 0.33048
N 4.53817 -1.75842 -0.15022
H 0.58336 -0.95225 0.31723
H 1.28163 0.28441 1.36605
H 0.27958 0.70318 -1.49720
H 1.06117 1.94551 -0.53634
H -0.99643 1.02651 1.19076
H -1.70248 0.82486 -0.42389
H -3.14641 2.28307 0.18112
H 2.51537 -0.38070 -1.35088
H 3.21581 0.56510 -0.03611
H 2.56983 -2.43772 0.05230
H 3.18710 -1.50773 1.42361
H 4.52020 -2.05523 -1.12541
H 4.97912 -2.51649 0.36884

metaopt_test_6.7.1_maxcycle20_output.txt
metaopt_test_6.7.1_maxcycle20_input.txt
metaopt_test_6.7.1_inertial_output.txt
metaopt_test_6.7.1_inertial_input.txt
metaopt_test_6.7.1_lbfgs_output.txt
metaopt_test_6.7.1_lbfgs_input.txt

[foxtran]

@PJpUnkie, you can try to play with #1233. I made some changes in this part, but not sure that all necessary changes were dome.