Different single-point GFN-FF energy with and without argument --verbose

[dargen3]

Hello,
I'd like to report a potential bug. TOTAL ENERGY for the same structure differs with and without the --verbose flag. The version of xtb is 6.6.0.

To Reproduce

P85371.zip

xtb P85371.pdb --gfnff --alpb water
results in TOTAL ENERGY = -35.183991089038 Eh

xtb P85371.pdb --gfnff --alpb water --verbose
results in TOTAL ENERGY = -35.353063192998 Eh

In both cases, I also get the warning but without an idea, of how to handle it and whether is it related to the error described above.
[WARNING] Runtime exception occurred
-1- gfnff_setup: Could not read topology file.

If the error is on my part, please guide me.
Thank you,
Ondřej

[Albkat]

Hi, thank you for reporting this.
It seems to be a bug.

[dargen3]

Thank you for your prompt reply. Do you have any idea which energy is correct, please?

[Albkat]

So the energy that comes from the SCC routine: TOTAL ENERGY = -35.183991089038 Eh, which is most likely the right energy.
But I recommend waiting until this issue is debugged.

[dargen3]

Hello, i have bad news. The bug is also in the 6.5.1 version. And it affects also structure optimization.

xtb P85371.pdb --gfnff --alpb water --opt
xtb P85371.pdb --gfnff --alpb water --opt --verbose
optimized_structures.zip

I know that it is pretty difficult to say, but is there any estimation, of when this issue can be debugged, please? Week, month, half of the year? I'm sorry to be in such a hurry, but this is quite critical for me and I have to decide what to do next. I'd try to look into it myself, but I'm completely out of Fortran.

[MtoLStoN]

Thank you for reporting this Issue. The difference in the output stems from here:

xtb/src/gfnff/gfnff_ini.f90

Lines 918 to 931 in 27203a8

	if(pr)then
	write(env%unit,'(''Hueckel system :'',i3,'' charge : '',i3,'' ndim/Nel :'',2i5, &
	& 3x, ''eps(HOMO/LUMO)'',2f12.6)')pis,ipis(pis),npi,nelpi,pisip(pis),pisea(pis)
	if(pisip(pis).gt.0.40) then
	write(env%unit,*)'WARNING: probably wrong pi occupation. Second attempt with Nel=Nel-1!'
	do i=1,mol%n
	if(piadr4(i).ne.0) write(env%unit,*) 'at,nb,topo%hyb,Npiel:', i,mol%sym(i),topo%nb(20,i),topo%hyb(i),piel(i)
	enddo
	nelpi=nelpi-1
	Api = Apisave
	call gfnffqmsolve(.false.,Api,S,.false.,300.0d0,npi,0,nelpi,dum,occ,eps) !diagonalize
	call PREIG(6,occ,1.0d0,eps,1,npi)
	endif
	endif

These routines are only activated when pr is set to true, which is essentially the same as setting the verbosity level with set%verbose. If these lines are commented out, the energy output will be the same whether or not the --verbose option is used, as it should be.

At the moment, I am not entirely confident with the force field code and cannot immediately identify the error (and it's late right now). I am also moderately confused that there seems to be a diagonalization that is only invoked with a verbosity option. @Thomas3R, do you have any idea?

[MtoLStoN]

So, I just checked with @sespic, and this actually seems to be a bug, that thankfully only hits a few specific pi systems. I will prepare a PR, that fixes this.

@dargen3 To answer your question, the energies and geometries, that are obtained with the --verbose should be the ones you would obtain without this bug. There seems to be some confusion here, sorry for that - we are still investigating.

[dargen3]

Thank you for the message and for working on the bug!

[dargen3]

Hello, thank you for resolving the error. I want to ask, how do I calculate for the correct energy? With or without the --verbose argument? Or are both options affected and I have to wait for the next version?

[MtoLStoN]

Hi @dargen3,
Both versions were affected. The error was fixed in #836. You can use xTB 6.6.1, which was released today.