Problem in relaxing cell dimensions when using hybrid functi

[jcconesa]

Hi,
I have found that when using GPUs for relaxing cell dimensions (ISIF=3) in calculations of paramagnetic lanthanide oxides involving a hybrid functional (e.g. HSE06) the results are weird: the cell volume increases by a factor of 3 or more, and the energy computed (probably correctly) in the first cycle rockets to much higher values in subsequent relaxation circles.
Has someone found a similar thing? I am using VASP 5.4.4, compiled with intel_parallel_studio_xe_2017, impi-2017.2.174 and cuda/9.2. The problem appears with both K80 and P100 GPUs. I am using these flags in INCAR:

ALGO=All; TIME=0.1 (Algo=Damped behaves similarly, and other TIME values make no change)
ISYM=0; LREAL=.TRUE.; LSCAAWARE = .FALSE. (other values except LREAL=.FALSE. behave similarly)
LASPH = .TRUE.; ADDGRID = .TRUE.; LMAXMIX = 6
PREC = Accurate; ENCUT=550; EDIFF=1E-5
IBRION = 2 (values of 1 and 3, with any POTIM, behave similarly)
NELM values are high, and MAGMOM values ensure that initial spin symmetry is broken.

The problem appears only, as I said, when GPUs, a hybrid functional and cell dimensions relaxation all occur simultaneously.
If this were a real bug in VASP it would be good to correct it before the new VASP 6 version were released.
José C. Conesa
Institute of Catalysis, CSIC, Madrid

[guiyang_huang1]

I also observed this problem.

It only happened for hybrid DFT and ISIF=3.

Using ISIF=4, and changing the lattice constant does not have problem.