MP2 ground state calculation - Tutorial: Difference between revisions

From VASP Wiki
No edit summary
No edit summary
Line 5: Line 5:
* '''MP2'''<ref name="marsman"/>: this was the first implementation in VASP and is very efficient for small systems, i.e. systems with less than ~40 valence electrons per unit cell or unit cells smaller than ~200 Å&#179;. The system size scaling of this algorithm is N&#8309;.
* '''MP2'''<ref name="marsman"/>: this was the first implementation in VASP and is very efficient for small systems, i.e. systems with less than ~40 valence electrons per unit cell or unit cells smaller than ~200 Å&#179;. The system size scaling of this algorithm is N&#8309;.
* '''LTMP2'''<ref name="schaefer2017"/>: the Laplace transformed MP2 algorithm has a lower scaling (N&#8308;) than the previous MP2 algorithm and is therefore efficient for large systems with unit cells larger than ~200 Å&#179; and more than ~40 valence electrons.
* '''LTMP2'''<ref name="schaefer2017"/>: the Laplace transformed MP2 algorithm has a lower scaling (N&#8308;) than the previous MP2 algorithm and is therefore efficient for large systems with unit cells larger than ~200 Å&#179; and more than ~40 valence electrons.
* '''stochastic MP2'''<ref name="schaefer2018"/>: faster calculations at the price of statistical noise can be achieved with the stochastic MP2 algorithm. It is an optimal choice for very large systems where only relative errors per valence electron (say 1 meV per valence electron) are relevant. Keeping the absolute error fixed, the algorithm exhibits a cubic scaling with the system size, N&#179;, whereas for a fixed relative errror, a linear scaling, N&#185;, can be achieved. Note that there is no k-point sampling and no spin polarization implemented for this algorithm.
* '''stochastic LTMP2'''<ref name="schaefer2018"/>: faster calculations at the price of statistical noise can be achieved with the stochastic MP2 algorithm. It is an optimal choice for very large systems where only relative errors per valence electron (say 1 meV per valence electron) are relevant. Keeping the absolute error fixed, the algorithm exhibits a cubic scaling with the system size, N&#179;, whereas for a fixed relative error, a linear scaling, N&#185;, can be achieved. Note that there is no k-point sampling and no spin polarization implemented for this algorithm.


At first, one should select the best algorithm according to the considered system size. In the following, a step by step instruction is presented.
Both ''LTMP2'' as well as ''stochastic LTMP2'' are high performance algorithms that can parallelize the MP2 calculation over thousands of CPUs.
 
At first, one should select the best algorithm according to the considered system size. In the following, a step by step instruction for each algorithm is presented.
 
<div style="background:#FFFFFF; border:solid 1px #AAAAAA; padding:0.5em 1em; margin-right:14em;">
# [[Preparation: the HF ground state|Preparation: the HF ground state]]
# [[Calculating the unoccupied HF orbitals|Calculating the unoccupied HF orbitals]]
# [[Using MP2|Using MP2]]
# [[Using LTMP2|Using LTMP2]]
# [[Using stochastic LTMP2|Using stochastic LTMP2]]
 
</div>


== References ==
== References ==

Revision as of 14:10, 27 March 2018

(UNDER CONSTRUCTION)

This tutorial introduces how to calculate the ground state energy using second order Møller-Plesset perturbation theory (MP2) with VASP. Currently there are three implementations available:

  • MP2[1]: this was the first implementation in VASP and is very efficient for small systems, i.e. systems with less than ~40 valence electrons per unit cell or unit cells smaller than ~200 ų. The system size scaling of this algorithm is N⁵.
  • LTMP2[2]: the Laplace transformed MP2 algorithm has a lower scaling (N⁴) than the previous MP2 algorithm and is therefore efficient for large systems with unit cells larger than ~200 ų and more than ~40 valence electrons.
  • stochastic LTMP2[3]: faster calculations at the price of statistical noise can be achieved with the stochastic MP2 algorithm. It is an optimal choice for very large systems where only relative errors per valence electron (say 1 meV per valence electron) are relevant. Keeping the absolute error fixed, the algorithm exhibits a cubic scaling with the system size, N³, whereas for a fixed relative error, a linear scaling, N¹, can be achieved. Note that there is no k-point sampling and no spin polarization implemented for this algorithm.

Both LTMP2 as well as stochastic LTMP2 are high performance algorithms that can parallelize the MP2 calculation over thousands of CPUs.

At first, one should select the best algorithm according to the considered system size. In the following, a step by step instruction for each algorithm is presented.

References