Linear dependent basis set for GFN1-xTB under PBC

[aronwalsh]

Describe the bug I have been having fun playing with xtb, but I keep running into a problem for crystals. I raise it here as I have checked with the conda version on mac, the precompiled linux binary, and a freshly compiled linux binary with ifort. It is possible that I am doing something silly as others don't seem to have the issue.

To Reproduce The diamond example runs fine with xtb --gfn 1 diamond, but for other cases I have checked including the 2-atom rocksalt MgO example, I get a fatal error related to diagnonalisation. For example, using the input file from the documentation:

$coord frac
    0.00      0.00      0.00      mg
    0.50      0.50      0.50      o
$periodic 3
$cell
 5.798338236 5.798338236 5.798338236 60. 60. 60.
$end

I get:

iter      E             dE          RMSdq      gap      omega  full diag
########################################################################
[ERROR] Program stopped due to fatal error
-6- Single point calculation terminated
-5- xtb_calculator_singlepoint: Electronic structure method terminated
-4- scf: Self consistent charge iterator terminated
-3- scc_core: Diagonalization of Hamiltonian failed
-2- mctc_lapack_sygvd: Failed to solve eigenvalue problem
-1- mctc_lapack_potrf: Factorisation of matrix failed
########################################################################

Running with xtb --gfn 1 --define --verbose confirms that the input geometry is read correctly and sensible.

Additional context Thanks for any help

[awvwgk]

The system is linear dependent and the overlap cannot be factorized by potrf anymore, therefore the calculation is stopped. Adding a cutoff for small eigenvalues in the overlap matrix would solve this issue.

[awvwgk]

It is not really a bug, but the currently expected behaviour with the very diffuse basis functions used for the xTB Hamiltonians. To fix it we need a robust way to project out linear dependencies from the tight binding basis set, which is not yet implemented for GFN1-xTB but already around in a somewhat instable way for GFN0-xTB.

[aronwalsh]

Thanks for the explanation @awvwgk. I thought I may be missing a trick. GFN0-xTB does indeed work for my examples. It could be worth highlighting in the PBC documentation.

[awvwgk]

The alternative is to finish the PBC implementations and make it finally a real thing, maybe we will get there this year.

[lgigli]

Hi, I am also observing this same issue, when using the GFN1-TB single point calculator (in pbc). The code snippet is the following:

from ase.io import read,write
from xtb.ase.calculator import XTB
xtb = XTB(method="GFN1-xTB")
f = read('LiPSdataset.xyz', '40')
f.calc = xtb
f.get_potential_energy()

I get:

XTBException: Single point calculation failed:
-5- xtb_calculator_singlepoint: Electronic structure method terminated
-4- scf: Self consistent charge iterator terminated
-3- scc_core: Diagonalization of Hamiltonian failed
-2- mctc_lapack_sygvd: Failed to solve eigenvalue problem
-1- mctc_lapack_potrf: Factorisation of matrix failed

Setting the pbc flag to False solves the issue.
How can one add a cutoff for small eigenvalues in the overlap matrix when dealing with PBC? I do not see any such option when defining the calculator, as the supported keywords are:

 method                   "GFN2-xTB"   Underlying method for energy and forces
 accuracy                 1.0          Numerical accuracy of the calculation
 electronic_temperature   300.0        Electronic temperatur for TB methods
 max_iterations           250          Iterations for self-consistent evaluation
 solvent                  "none"       GBSA implicit solvent model
 cache_api                True         Reuse generate API objects (recommended) 

Thanks for any help on this.