Convergence in liquid metals

[gpanchap]

In my liquid Si runs, as soon as I include empty states the total energy doesn't seem to converge anymore. This is my input: (with nempty=0 the total energy converges ~mRy in about 200 scf. steps):

si2.sys #unit cell, psp and positions for N=200 atoms
set ecut 15 #kin. en. cutoff in Ry
set empty 50 #Total of 450 states
set fermi_temp 1000 #temp. in K. (Entropy=wf_entropy*kB*fermi_T/2.0)
set wf_dyn PSD
set charge_mix_coeff 0.001 #small value for metals
set ecutprec 5
run 0 1000

The energy appears to converge till the occupation gets smeared, which is then followed by a large jump in the energy (~ 200Ry) and rapid oscillations in energy after that. Interestingly, after a few hundred steps the entropy goes to zero (no occ. smearing), and the oscillation is between 'two' limiting values, each of which is well converged.

Last few lines of etotal_int:

<etotal_int> -547.09271275 </etotal_int>
<etotal_int> -548.52868450 </etotal_int>
<etotal_int> -547.09364208 </etotal_int>
<etotal_int> -548.52820650 </etotal_int>
<etotal_int> -547.09454015 </etotal_int>
<etotal_int> -548.52761111 </etotal_int>
<etotal_int> -547.09549035 </etotal_int>
<etotal_int> -548.52715518 </etotal_int>


Would you know what could be going on, and if there is a better scheme to achieve convergence in metals ?

[gpanchap]

Including a non-zero value for nite and changing the mixing coefficient to 0.8 gets rid of this problem. But convergence is still quite slow, requiring 'run 0 50 5' initially to converge to ~ 10mRy. Subsequently it requires 'run 100 30 5' . Is it possible to enter a tolerance in convergence (energy, wfk, etc.) so that I don't have to perform the same number of steps (i.e nscf*nite) at each ionic step? Thanks.

[fgygi]

This feature is not yet available. The implementation of a convergence criterion for the termination of the self-consistent loop is being considered in upcoming releases.