Si He core upf

[JMuff22]

Could a .upf be added for Si with He core? Is there a reason this has not already been included?

Only asking because having trouble using ppconvert to convert the pseudopotentials.

[mcbennet]

Hi Jake, the smaller core PPs can be difficult to converge -- this is probably why the .upf files have not been populated for this set of potentials. However, I just tried Si and (luckily) it converged without issue.

I have attached it here.

si-he-core_share.zip

[JMuff22]

Thanks for this! Why are they difficult to converge?

How did you generate these files? Would be very interested to know!

[prckent]

@mcbennet in this case is the convergence tricky with the atomic solver in ppconvert needed to make the projectors or is the convergence you are referring to the plane wave cutoff and self-consistency in an eventual QE calculation?

[mcbennet]

Generally, for the smaller cores, the atomic solver can be the difficult hurdle, yes. However, for Si I was able to get over this by playing with the cutoff radii and choice of local channel. The transferability can also be poor for these as well -- agreement between semilocal and KB -- Si was below 0.1eV errors for a number of states, so I considered it acceptable. Lastly, indeed the PW cutoff can be pretty high, where Si here looks to be around 900Ry.

Gani is looking at the 2nd row now and attempting to fill in the missing UPFs. From his initial findings, some look to have unreasonable PW cutoffs (we're attempting to confirm this) and unreasonable transferability. If the latter, then these edge cases will likely be excluded from any future PR until their transferability is resolved.

@JMuff22 we are using a modified opium code to generate these projectors.

[camelto2]

@mcbennet in this case is the convergence tricky with the atomic solver in ppconvert needed to make the projectors or is the convergence you are referring to the plane wave cutoff and self-consistency in an eventual QE calculation?

Just to chime in here, most of our conversions have come from the modified version of opium. This is because it has some extra checks that ppconvert doesn't have (atomic state testing, checking PW cutoffs, ghost state testing at the DFT level). However, in a lot of cases I have found that opium has some trouble with the atomic solver. In a case where opium doesn't converge, I've noticed that ppconvert generates the UPF problem with no problem. The problem is then that we don't have all of those same checks in ppconvert. I plan to look into adding these and add spin-orbit since the solver seems to be more robust than opium

[prckent]

Why not improve the solver in opium?

[camelto2]

Also an option, but I haven't had a chance to look in detail for why it breaks when it does. But I also think having an alternative path (especially for spin-orbit) to do cross validation is useful. Some of the checks that are missing from ppconvert (ghost states, cutoff checking) are straightforward to add

[prckent]

Good to know. It does need some serious cleaning though - while it appears to run OK and pass our sanity check tests, the code sanitizers don't like it since it touches plenty of uninitialized memory.

[JMuff22]

Unsure how relevant this if but I've found that using ppconvert to convert UPF to casino PP to not be that good. The number of grid points goes from 1144 to 10001 and the values given are different to the UPF PP mesh.

Any tips on using ppconvert for upf/gamess to casino?

[prckent]

@JMuff22 just merged Gani's PR with the extra potentials. Please let us know how you get on with them.

[JMuff22]

@prckent We didn't find much success with the Si He core PPs. We were trying to test them out through calculating the cohesion energy of Si using the CASINO QMC program. Unsure if it was due to the conversion of the upf PP to casino format or something in the input files but we found that the cohesion energy was significantly off theoretical values.

[prckent]

Thanks for the update Jake. I am sure everyone here will be interested in the conclusions that you eventually draw. Presumably you have seen https://doi.org/10.1103/PhysRevB.103.205206