I am working on Pt(111) / nanoparticles with various ligands, I am mainly using two softwares:

  • A well known general purpose plane-wave code (Quantum Espresso).
  • A less known software that is also aiming general purpose with plane-wave accuracy, the theory is a little bit more complicated (Wannier functions). It has been actively developed for 20 years.

The second software has difficulties performing geometry optimisations to a force threshold that I would consider not really tight (0.25 eV/A). The force gradient has an oscillating behaviour, it is getting down but really slowly (50 geometry iterations to get to 0.4 eV/A) as a comparison Quantum Espresso get to 0.01 eV/A in 30 iterations.

From what I have seen so far this problem seems to be a general problem and not only restricted to my system.

I have different persons telling me two possible explanations:

  • This problem is inherent to Wannier wavefunctions DFT, due to the truncation applied on the orbitals, forces will have some noise.
  • This problem is due to a bad geometry optimiser which needs to be rewritten.

Is there anyone on this Stack Exchange with experience on this topic or any references that might help my case?

  • 1
    $\begingroup$ Can you name the second software? $\endgroup$ Mar 23, 2022 at 15:41
  • $\begingroup$ @TristanMaxson I am afraid I can't, if this makes the question too general/hard to answer I will delete it. $\endgroup$
    – Okano
    Mar 23, 2022 at 15:58
  • $\begingroup$ @Okano I don't know if it would be answerable without knowing the software. The details of how the optimizer works are likely highly dependent on the program and any number of different factors/parameters could cause problems in specific cases. I suppose a question about the first bullet (are Wannier functions susceptible to noise in computed forces?) would be workable without knowing the software. But if you want an answer that relates to a specific program, I think it will be impossible without knowing what that program is. $\endgroup$
    – Tyberius
    Mar 23, 2022 at 16:23
  • 2
    $\begingroup$ I have met this problem before with NGWF methods. I strongly recommend determining the noise in the forces (ideally with a numerical derivative of the energy with respect to atomic positions) and then increasing the truncation radius of the NGWFs and repeating. Standard BFGS/LBFGS methods are not very tolerant of noise. Are you optimising the NGWFs or just the density kernel? I'd be surprised if the geometry optimiser is as poor as this, but if it has an ASE interface, you could try its geometry optimisation methods. $\endgroup$ Mar 23, 2022 at 23:15
  • 1
    $\begingroup$ Any update on this issue? If you'd like to drop me an email, I can look at this in more detail and we can post the answer here when we've resolved it. $\endgroup$ Apr 7, 2022 at 0:21