IPEA hydrogen basis overcomplete

[bhourahine]

Describe the bug
Running a structure consisting of multiple H2 molecules gives NaN from the eigensolver. The overlap matrix seems to become non-positive definite for larger systems, suggesting the H basis is over-complete (possibly too diffuse s*).

To Reproduce
Steps to reproduce the behaviour:

See for xtb version 6.7.1 (7618f60), self-compiled, but seems to be from the parameterisation, not specific code version. Geometry available here.

xtb --gfn 1 --vip structure.xyz

Self-consistency cycle looks like:

iter E dE RMSdq gap omega full diag
1 NaN NaN NaN NaN 0.0 T
2 NaN NaN NaN NaN 100000.0 T
.
.
.

Expected behaviour

Run to self-consistency successfully.

[awvwgk]

Thanks for the report, Ben. The IPEA parametrization does indeed add diffuse s functions for H and the second row elements.

One way to remove the linear dependencies is to project those functions out or to reparametrize the IPEA version of GFN1.

[bhourahine]

Yes, reparameterization would probably be better (possibly s* is too diffuse, so they become dependent once there are many atoms). Projection is usually not such a good idea (looking at the spectrum of the overlap matrix, this is actually going negative in some structures, not just the usual over complete basis null values).