About the max_forces and max_stress in GeometryOpt.

[fangyongxinxi]

Dear,

I want to get the acceptable parameters for my system, so I did a loop calculation like below:
 
for ecut/mas_force/k-points in (a1,a2,a3,a4,....):
      do calculation
the ecut and k-points are OK for me, but the max_stress or max_forces, always bothers me. Take max_stress for example, in my test calculation, the results are like this:

mas_stress(eV/Ang**3)       delta(E/eV)
0.2   -1.269
0.18   -1.269
0.15   -1.269
0.1   -1.269
0.08   -1.269
0.06   -1.269
0.05   -1.8165
0.03   -2.11
0.02   -2.11
0.015   -2.187
0.013   -2.187
0.008   -2.201
0.006   -2.201
0.002   -2.204
0.001   -2.204

the results are step-like drop, seems not convergent. The definition of stress can be found in the manual, and the default value for max_stress is 0.05.

So, my question is: could you give me some experience on how to set the max_stress, or which one is acceptable.

Thanks~

[zh]

1 eV/Angstrom^3 = 160.2176487 GPa,  Typically, several GPa may be acceptable for the maximum stress. So 0.05 eV/Angstrom^3 in the default setting is suitable.

In your test case,  0.03 eV/Angstrom^3 for the maximum stress gives a relatively high accuracy.

[Anders Blom]

It's a bit unusual to do a loop for the max_stress, it's more common to do the stress over the lattice constant, and in that case the curve will be smooth. The reason it's step like is simply because depending on the max_stress the SCF loop will terminate on different steps. So it cannot be smooth - if e.g. after step 15 the stress is 0.04 and in step 16 it is 0.01, then for any value between 0.01 and 0.04 it will be considered converged at step 15.

Some systems have a very weak total energy/stress dependence with lattice constant, some a large. For bulk metals and perhaps oxides it is true that the default or a bit below is accurate, but on the other hand I have seen that for some materials with lower dimensionality, like nanotubes or planar structures, you may need much lower values if you really want an accurate lattice constant.