Relaxation of a big supercell with the muon is a big bottleneck... so we should consider some strategy and pre-processing to speed up calculations. I have already some arguments ...

to be refined. I have some in the good notes, and also:

Additional Optimize convergnece of muons:

(1) if metal: muonium —> faster to converge : pre process step to evaluate the gap?
(2) if gbrv: low cutoffs —> faster the calculations and they converge faster anyway
(3) if ihighly inohomogeneous —> local TF: faster to converge —> pre process step in which we analyse this? should be related to the # of symmetries, but without the muon… otherwise it is always no symmetry. mixing_mode = 'local-TF'

(4) k points more than only gamma seems to be needed to converge better, in particular with the U.
(5) for the magnetic inequivalents to be relaxed agains, maybe you can restart with the charge density of the previous? not so sure, the muon position is changing.

seems that the local mixing helps more than 0.3 for convergence.
m1 : local TF and mixing 0.7
m2: no local TF and mixing 0.3
m3: local TF and mixing 0.3 seems a little bit slower than m1 (but you should try m1 with the m3 settings to be sure… same for m2) — m1 and m3 took the same number of iterations actually to reach the first convergence… see for the forces, how many iteration before full relaxation they took: 27 mins for m3, 22 mins for m1 :)

m4: no local TF and mixing 0.7 —> sooo bad !! THE LOCAL TF IS THE REAL PLAYER HERE!!!!! (and I guess also the low cutoffs and muonium).

From Pietro
Let me elaborate a little. Two points are important:

1. experimental magnetic moment in Fe is 0.6 mu_B, I think DFT gives ~ 2
   mu_B. This should not be a problem since we use the magnetic moments in
   the mcif.

2. There are many equilibrium positions that average out. There should
   be 4 above/below the Fe plane and the same above/below the O plane.

Two points to keep in mind for the future:

1. force convergence should be small and it's not unexpected to require
   100s steps for convergence

2. maybe gamma is a little too small, but this you already said.

Also, some pseudo can help

This tries to tackle one problem of the relax scf convergence
The more the symmetry, the more convenient the conventional cell instead of primitive (see lafeaso).

• to understand if this is related to the cubic symmetry or just the number of symmetries. I think is the cubic one.
• we need in some way an interactive decision on the cell, or a button before the upload, as I have implemented. But for now it is not so good/intuitive.

• We want to introduce the GBRV family of pseudopotentials, which are
  commonly used for Muon simulation and lead to smoother convergences in the relaxations of muon sites. This was verified in LaFeAsO, problem was La.

• it remains the problem that the pseudos cannot be upload in the app via advance tab, as we cannot upload the H one (it is introduced on the fly)

• produce a sort of gbrv pseudo family and pseudo family cutoff to be delivered in some way online, so easily accessible

• produce it in such a way to be used in QE workchains protocol as pseudo_family. for now not working as this. so we first create protocol with the default pseudo family (SSSP...) then on the fly in the workchain we update the pseudos. NB: this is required also because we insert the muon (H) later, during the run.

I think the slow part is the KT function.