Bulk silver - lattice constant calculations

[physicalattraction]

I have heard that in order to trust your results, they should (at least) converge for varying k-meshes and varying cutoff energies. I have tried to calculate the lattice constant (experimental: a = 4.08 A) in two ways, each with 4 different k-meshes (8x8x8 to 11x11x11) and 4 different cutoff energies (250 - 400 eV), and each with the PW91 XC functionals with the PAW method.

The first method is to build a script with a loop to vary the lattice constant in POSCAR and without any ionic updates in the INCAR, similar to the example in 9.2 of the VASP manual. I find a = 4.16 A in 15 of the 16 calculations. F = -2.73 eV in 15 of the 16 calculations.

The second method is to start with a = 4.16 in the POSCAR and add to INCAR: NSW = 10, IBRION = 2, ISIF = 3 to allow for unit cell shape and size changes. This results in the following optimal lattice constants:
ENMAX | optimal a
250 | 4.13
300 | 4.14 - 4.15
350 | 4.15
400 | 4.15
Again, F = -2.73 eV in 15 of the 16 calculations.

The question now is: how reliable are these results? I know there is no unambiguous answer to this questions, but I wonder if from this information one can deduce that the optimal a = 4.15-4.16 A, how crucial the exact lattice constant is for other physical quantities (i.e. how many decimals are significant?) and if it is not a problem that this result differs from the experimental result.

[admin]

1) for automatic optimization of the cell using stress tensors (ISIF>0), you need to
-- increase the cutoff to at least 1.3*ENMAX (which would be 325 eV for paw-PW91 Ag). Please have a look for the ISIF tag in the online manual --> the results of 250 and 300 eV are not accurate
-- PREC should be set to accurate
-- the basis set is not changed during one vasp run. Therefore, if you start your automatic optimization with a lattice volume which is far away from the equilibrium volume, you will have considerable basis set errors. Please do 2-3 optimization steps, then start from the CONTCAR of that run without using WAVECAR. If the volume/cell shape changes of this run still are considerable, please repeat this procedure.

-- if you require very high accuracy, it is recommended to do an automatic optimization first, and do a conventional E/V fit of the parabola next to the minimum (a few single step runs)

-- concerning the general errors wich are made by DFT calculations (LDA overbinds, PW91 underbinds): this behaviour is well documented in literature.