How should I rattle an ASE Atoms object to improve convergence in DFTB+?

ASE
1

Python script (snippet):
from ase import io
atoms = io.read(“structure.xyz”)

Tried several stdev values [0.0001 to 0.1]

atoms.rattle(stdev=0.0001, seed=None)
io.write(“structure2.xyz”, atoms)

DFTB+ input file (snippet):
Driver = GeometryOptimization {
LatticeOpt = Yes
FixAngles = Yes
Optimizer = Rational {}
MaxSteps = 1000
Convergence {
Energy = .02
GradElem = 0.000388
}
}

Hamiltonian = xTB {
Method = “GFN1-xTB”
MaxSccIterations = 1000
}

No matter what stdev I use between 0.0001–0.1, the geometry doesn’t converge. Is rattle() even the right tool here, or is there a better way to slightly rattle structures for DFTB+ convergence? I don’t want to use rng=numpy.random due to replicability.

DFTB+ convergence errors:
“SCC is NOT converged, maximal SCC iterations exceeded”
“Failure in diagonalisation routine ZHEEVR, unable to complete Cholesky factorization”
“Fermi level search did not converge”

2

Rattling doesn’t fix every possible convergence problem! This looks like more of a DFTB+ problem then an ASE issue.

I have run into stability problems with GFN1-xTB sometimes.
You could try using one of the other electronic solvers in the Hamiltonian section, such as DivideAndConquer, and/or a different “Filling” method.