I am currently trying to run RPA calculations on some platinum clusters. For this I am using ridft to us the resulting PBE orbitals for RPA, using a def2-QZVPP basis. I can successfully do this through a combination of heavy damping, many iterations, orbital shifts, etc. However, when these converge, it results in a negative HOMO-LUMO gap. This is a problem for performing RPA calculations, as it requires an aufbau population (cf.
https://forum.turbomole.org/index.php?topic=531.0).
A solution to this was to use the $lastdiag keyword to force a last diagonalisation and obtain an aufbau population. This works well for unrestricted calculations and the subsequent RPA calculations are easy to perform. However, for restricted calculations, the use of $lastdiag results in a large, positive HOMO-LUMO gaps (~7 eV), which makes the RPA energies unphysical. I am using Turbomole 7.3.1 but found this issue across multiple versions
The problem does is not with the RPA, or DFT convergence, as the RPA works well for unrestricted calculations, and the DFT energies are reasonable. The issue seems to be with using $lastdiag with restricted DFT. I know that this is a issue that has been found in the past (
https://forum.turbomole.org/index.php?topic=531.0) but found no solutions in the forum. Is there a way to solve this?
Finally, we tested several key words to see if they would have any effect on the large positive HOMO-LUMO gap (scfconv, scftol, denconv, dft gridsize, thime, thize, scforbitalshift, etc) but found that the only thing that made any difference was $scftol, the integral screening threshold. This was not consistent however and the HOMO-LUMO gap changed seemingly randomly, irrespective of $scftol:
$scftol / 10
-x HOMO-LUMO Gap/ eV
11 0.07
12 9.46
13 0.53
14 1.18
15 0.30
16 0.10
17 6.22
18 8.06
Do you have any insight into this or new suggestions that we could try?
Best wishes,
Chris