Hexagonal lutetium oxide with r2 SCAN and rSCAN

Dear Users,

I'm trying to run a regularized SCAN calculation on a hexagonal lutetium oxide, input below.

The problem is: it fails, no matter what I do. With r2 SCAN (xctype 100 497 498) it usually happens in 4 steps, with rSCAN (xctype 100 493 494) it can last, but then it fails.
I tried running RPP mGGA first (xctype 100 209 9), and then use that STATE.OUT to start the rSCAN calculations. It helps them to make more steps, but eventually they fail.

The fail is a rapid increase in the RMS, super high values of energies (1e36) and core d orbital energies (Lu 3d and 4d, 1e26-1e36), and the corresponding "Warning(rdirac): maximum iterations exceeded" and "Error(genapwfr): degenerate APW radial functions".

I tried different basis sizes (via rgkmax and gmaxvr) and, independently:
- highq
- autolinengy=t
- nxapwlo=1,2
- ptnucl=f
- non-default constant linearization energies put roughly at the valence band maximum (at least, this is what works with RPP)

Also, I wonder if metagga keyword is automatically set to true if mGGA is used.

Anyway, the basic input is below.
Would anyone know how to converge it?

Thank you in advance.
Andrew

tasks
  0

!metagga 
!  t

xctype
  100 497 498
!  100 493 494
!  100 209 9

autokpt
 t

! scale the number of radial mesh points in the muffin-tins
nrmtscf
 1.5

highq
 t

nwrite
 5

avec
   6.735554000       0.000000000       0.000000000    
  -3.367777000       5.833160873       0.000000000    
   0.000000000       0.000000000       10.90088900    

atoms
   2                                    : nspecies
'Lu.in'                                 : spfname
   2                                    : natoms; atposl, bfcmt below
    0.33330000    0.66670000    0.25040000    0.00000000  0.00000000  0.00000000
    0.66670000    0.33330000    0.74960000    0.00000000  0.00000000  0.00000000
'O.in'                                  : spfname
   3                                    : natoms; atposl, bfcmt below
    0.00000000    0.00000000    0.00000000    0.00000000  0.00000000  0.00000000
    0.33330000    0.66670000    0.64640000    0.00000000  0.00000000  0.00000000
    0.66670000    0.33330000    0.35360000    0.00000000  0.00000000  0.00000000

Hi Andrew,

I've managed to get regularized SCAN to converge with your crystal structure.

The trick was to first converge a non-relativistic calculation (solscf=100) and then restart with relativity included (solscf=1). Some smoothing of the numerical gradients was also needed. Here is the input file:

tasks
  0

msmooth
  4

solscf
  100.0

xctype
  100 493 494

ngridk
  4  4  4

avec
   6.735554000       0.000000000       0.000000000
  -3.367777000       5.833160873       0.000000000
   0.000000000       0.000000000       10.90088900

atoms
   2                                    : nspecies
'Lu.in'                                 : spfname
   2                                    : natoms; atposl, bfcmt below
    0.33330000    0.66670000    0.25040000    0.00000000  0.00000000  0.00000000
    0.66670000    0.33330000    0.74960000    0.00000000  0.00000000  0.00000000
'O.in'                                  : spfname
   3                                    : natoms; atposl, bfcmt below
    0.00000000    0.00000000    0.00000000    0.00000000  0.00000000  0.00000000
    0.33330000    0.66670000    0.64640000    0.00000000  0.00000000  0.00000000
    0.66670000    0.33330000    0.35360000    0.00000000  0.00000000  0.00000000

It will obviously need to be converged better, particularly with respect to the number of empty states. The option 'metagga' is not automatically set, it just switches some parameters to those recommended for meta-GGA calculations, like smoothing, empty states and the number of radial points.

Regards,
Kay.

https://sourceforge.net/p/elk/discussion/897820/thread/2239dce205/