Stochastic LTMP2

From VASP Wiki

On this page, we explain how to perform a calculation using the stochastic LTMP2[1] algorithm. Make sure you successfully completed the preparation steps Hartree-Fock ground state and Hartree-Fock virtuals.

NOTE: If you use this algorithm, please cite reference [1] in your publication in addition to the standard VASP reference.

The INCAR file

The LTMP2 calculation can simply be performed using the following INCAR file

ALGO = ACFDTRK 
LSMP2LT = .TRUE.
ESTOP = # accuracy of energy per tau-point
NSTORB = # number of stochastic orbitals per cycle
NOMEGA = # number of tau-points (see below)
NBANDS = # same valule as in the Hartree-Fock unoccupieds step ( = number of plane-waves)
ENCUT = # same value as in the Hartree-Fock step
LORBITALREAL = .TRUE.
PRECFOCK = Fast

Make sure that VASP reads the WAVECAR file from the Hartree-Fock virtuals step. The setting for PRECFOCK is strongly recommended, since the code heavily relies on real space grid FFTs.

NOMEGA flag

The number of -points is controlled by the NOMEGA flag. This is necessary to calculate the Laplace transformed energy denominator (see Ref [1] for details),

Usually it is sufficient to set NOMEGA to 6. For materials with a small bandgap it is worth checking if the MP2 energy changes with increasing NOMEGA (e.g. 8 or 10). Note, that the MP2 energy diverges with 1/bandgap, independent of NOMEGA.

ESTOP flag

This flag defines the stop condition for the stochastic algorithm. It defines the energy accuracy in units of eV for each individual tau-point of the two individual MP2 energy contributions (direct MP2 term + exchange MP2 term). Since the statistical errors of each contribution is independent, the standard deviation of the MP2 energy can be estimated as

According to our experience, the error of the resulting MP2 energy can then be safely estimated by .

Thus, if you require an MP2 energy with a maximum error of , you should set

NSTORB flag

This flag defines the number of stochastic orbitals per cycle, i.e. the number of stochastic orbitals that define one stochastic sample. If the sample is not large enough, the calculations is repeated until the accuracy, defined by ESTOP, is reached.

As a rule of thumb we recommend to set

Parallelization

The stochastic LTMP2 algorithm supports parallelization with MPI and OpenMP (OMP). The optimal setting is to set the number of MPI ranks as well as the KPAR flag to the number of cores (#cores), i.e. start VASP using

mpirun -np #cores vasp

and write

KPAR = #cores

in the INCAR file. With this setting the entire set of Hartree-Fock orbitals (WAVECAR) is available on each MPI rank, which is necessary to calculate stochastic samples independently. Note, that KPAR is only used to control the distribution of the orbitals and has nothing to do with k-point parallelization here.

However, for very large systems (large WAVECAR files) the available storage per MPI rank could be insufficient to store the entire set of orbitals. In this case, simply decrease the KPAR. Note that the available memory for the orbitals can be calculated by (memory per MPI rank) * (number of MPI ranks) / KPAR. For example, if your WAVECAR file has 17 GB you need 2*17 GB = 34 GB of memory to distribute the orbitals (the factor 2 is due to double precision). If want to use 64 cores with 4 GB per core and 64 MPI ranks, you have to set KPAR = 4. In this case the orbitals are distributed over 64/4 = 16 MPI ranks. Each MPI rank will still be able to perform independent stochastic calculations, however, a bit more MPI communication is necessary.

It is also possible to increase the memory per MPI rank using shared memory with OMP. This is a viable option if your available memory per core is too small, decreasing KPAR does not help or you don't want to set too small KPAR values. However, in general, it is recommended to solve memory issues with the KPAR flag first.

References