scnc
/
qmcpack
mirror of https://github.com/QMCPACK/qmcpack.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

manual: update environments and references to Nexus in the condensed matter lab 

git-svn-id: https://subversion.assembla.com/svn/qmcdev/trunk@6778 e5b18d87-469d-4833-9cc0-8cdfa06e9491
This commit is contained in:
Jaron Krogel 2016-02-16 21:16:35 +00:00
parent a4e9f631db
commit d0d0ff8590
1 changed files with 29 additions and 17 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

46
manual/lab_condensed_matter.tex
View File

@ -13,7 +13,7 @@ DE-AC04-94AL85000.}
The goal of this lab will be to introduce you to the somewhat specialized problems involved in performing diffusion Monte Carlo calculations on condensed matter as opposed to the atoms and molecules that were the focus of earlier labs. Calculations will be performed on two different systems. Firstly, we will perform a series of calculations on BCC beryllium focusing on the necessary methodology to limit finite size effects. Secondly, we will perform calculations on graphene as an example of a system where qmcpack’s ability to handle cases with mixed periodic and open boundary conditions is useful. This example will also focus on strategies to limit memory usage for such systems.
All of the calculations performed in this lab will utilize the project suite tools that vastly simplify the process by automating the steps of generating trial wavefunctions and performing DMC calculations.
All of the calculations performed in this lab will utilize the Nexus workflow management system that vastly simplifies the process by automating the steps of generating trial wavefunctions and performing DMC calculations.
\newcommand{\vp}{\mathbf{a}^\text{p}}
\newcommand{\vs}{\mathbf{a}^\text{s}}
@ -73,11 +73,12 @@ In the case of a BCC crystal, it is possible to tile the one atom primitive cell
\end{array}\right]
\end{equation}
We will now use the project suite to generate the trial wavefunction for this BCC beryllium.
We will now use Nexus to generate the trial wavefunction for this BCC beryllium.
Fortunately, the project-suite will handle determination of the proper k-vectors given the tiling matrix. All that is needed is to place the tiling matrix in the Be-2at-setup.py file. Now the definition of the physical system is:
Fortunately, the Nexus will handle determination of the proper k-vectors given the tiling matrix. All that is needed is to place the tiling matrix in the Be-2at-setup.py file. Now the definition of the physical system is:
\begin{lstlisting}
\begin{shaded}
\begin{verbatim}
bcc_Be = generate_physical_system(
lattice = 'cubic',
cell = 'primitive',
@ -92,25 +93,31 @@ Fortunately, the project-suite will handle determination of the proper k-vectors
kgrid = kgrid,
kshift = (.5,.5,.5)
)
\end{lstlisting}
Where the tiling line should be replaced with the row major tiling matrix from above. This script file will now perform a converged DFT calculation to generate the charge density in a directory called bcc-beryllium/scf and perform a non self consistend DFT calculation to generate single particle orbitals in the direcotry bcc-beryllium/nscf. Fortunately, the project suite will calculate the required k-points needed to tile the wavefunction to the supercell, so all that is necessary is the granularity of the supercell twists and whether this grid is shifted from the origin. Once this is finished, it performs the conversion from pwscf's binary format to the hdf5 format used by qmcpack. Finally, it will optimize the coefficients of one-body and two-body jastrow factors in the supercell defined by the tiling matrix.
\end{verbatim}
\end{shaded}
Where the tiling line should be replaced with the row major tiling matrix from above. This script file will now perform a converged DFT calculation to generate the charge density in a directory called bcc-beryllium/scf and perform a non self consistend DFT calculation to generate single particle orbitals in the direcotry bcc-beryllium/nscf. Fortunately, Nexus will calculate the required k-points needed to tile the wavefunction to the supercell, so all that is necessary is the granularity of the supercell twists and whether this grid is shifted from the origin. Once this is finished, it performs the conversion from pwscf's binary format to the hdf5 format used by qmcpack. Finally, it will optimize the coefficients of one-body and two-body jastrow factors in the supercell defined by the tiling matrix.
Run these calculations by executing the script Be-2at-setup.py. You will notice that such small calcuations as are required to generate the wavefunction of Be in a one atom cell are rather inefficent to run on a high performance computer such as vesta in terms of the time spent doing calculations versus time waiting on the scheduler and booting compute nodes. One of the benefits of the portable hdf format that is used by qmcpack is that you can generate data like wavefunctions on a local workstation or other convenient resource and only use high performance clusters for the more expensive QMC calculations.
In this case, the wavefunction is generated in the directory bcc-beryllium/nscf-2at\_222/pwscf\_output in a file called pwscf.pwscf.h5. It can be useful for debugging purposes to be able to verify the contents of this file are what you expect. For instance, you can use the tool h5ls to check the geometry of the cell where the dft calculations were performed, or number of k-points or electrons in the calculation. This is done with the command: h5ls -d pwscf.pwscf.h5/supercell or h5ls -d pwscf.pwscf.h5/electrons.
In the course of running Be-2at-setup.py, you will get an error when attempting to perform the vmc and wavefunction optimization calculations. This is due to the fact that the wavefunction has been generated supercell twists of the form (+/- 1/4, +/- 1/4, +/- 1/4). In the case that the supercell twist contains only 0 or 1/2, it is possible to operate entirely with real arithmetic. The executabe that has been indicated in Be-2at-setup.py has been compiled for this case. Note that where this is possible, the memory usage is a factor of two less than the general case and the calculations are somewhat faster. However, it is often necessary to perform calculations away from these special twist angles in order to reduce finite size effects. To fix this, delete the directory bcc-beryllium/opt-2at, change the line in near the top of Be-2at-setup.py from
\begin{lstlisting}
\begin{shaded}
\begin{verbatim}
qmcpack = '/soft/applications/qmcpack/build_XL_real/bin/qmcapp'
\end{lstlisting}
\end{verbatim}
\end{shaded}
to
\begin{lstlisting}
\begin{shaded}
\begin{verbatim}
qmcpack = '/soft/applications/qmcpack/build_XL_complex/bin/qmcapp'
\end{lstlisting}
\end{verbatim}
\end{shaded}
and rerun the script.
When the optimiztion calculation has finished, check that everything as proceeded correctly by looking at the output in the opt-2at directory. Firstly, you can grep the output file for Delta to see if the cost function has indeed been decreasing during the optimization. You should find something like:
\begin{lstlisting}
\begin{shaded}
\begin{verbatim}
OldCost: 4.8789147e-02 NewCost: 4.0695360e-02 Delta Cost:-8.0937871e-03
OldCost: 3.8507795e-02 NewCost: 3.8338486e-02 Delta Cost:-1.6930674e-04
OldCost: 4.1079105e-02 NewCost: 4.0898345e-02 Delta Cost:-1.8076319e-04
@ -119,9 +126,11 @@ When the optimiztion calculation has finished, check that everything as proceede
OldCost: 4.2176276e-02 NewCost: 4.2083371e-02 Delta Cost:-9.2903058e-05
OldCost: 4.3977361e-02 NewCost: 4.2865751e-02 Delta Cost:-1.11161830-03
OldCost: 4.1420944e-02 NewCost: 4.0779569e-02 Delta Cost:-6.4137501e-04
\end{lstlisting}
\end{verbatim}
\end{shaded}
Which shows that the starting wavefunction was fairly good and that most of the optimizaiton occurred in the first step. Confirm this by using qmca to look at how the energy and variance changed over the course of the calculation with teh comand: qmca -q ev -e 10 *.scalar.dat executed in the opt-2at directory. You should get output like the following:
\begin{lstlisting}
\begin{shaded}
\begin{verbatim}
LocalEnergy Variance ratio
opt series 0 -2.159139 +/- 0.001897 0.047343 +/- 0.000758 0.0219
opt series 1 -2.163752 +/- 0.001305 0.039389 +/- 0.000666 0.0182
@ -131,7 +140,8 @@ opt series 4 -2.162441 +/- 0.000865 0.039597 +/- 0.000342 0.0183
opt series 5 -2.161287 +/- 0.000732 0.039954 +/- 0.000498 0.0185
opt series 6 -2.163458 +/- 0.000973 0.044431 +/- 0.003583 0.0205
opt series 7 -2.163495 +/- 0.001027 0.040783 +/- 0.000413 0.0189
\end{lstlisting}
\end{verbatim}
\end{shaded}
Now that the optimization has completed successfully, we can perform dmc calculations. The first goal of the calculations will be to try to eliminate the one body finite size effects by twist averaging. The script Be-2at-qmc.py has the necessary input. Note on line 42 two twist grids are specified, (2,2,2) and (3,3,3). Change the tiling matrix in this input file as in Be-2at-qmc.py and start the calculations. Note that this workflow takes advantage of qmcpack's ability to group jobs. If you look in the directory dmc-2at\_222 at the job submission script, (dmc.qsub.in) you will note that rather than operating on an xml input file, qmcapp is targeting a text file called dmc.in. This file is a simple text file that contains the names of the 8 xml input files needed for this job, one for each twist. When operated in this mode, qmcpack will use mpi groups to run multiple copies of itself within the same mpi context. This is often useful both in terms of organizing calculations and also for taking advantage of the large job sizes that computer centers often encourage.
@ -149,16 +159,18 @@ In this section we will examine a calculation of an isolated sheet of graphene.
Aside from the change in boundary conditions, the main thing that distinguished this kind of calculation from the beryllium example above is the large amount of vacuum in the cell. While this is a very small calculation designed to run quickly in the tutorial, in general a more converged calculation would quickly become memory limited on an architecture like BG/Q. When the initial wavefunciton optimizaiton has completed to your satisfaction, run the scripts graphene-loop-buffer.py and graphene-loop-mesh.py. These examine within variational Monte Carlo two approaches to reducing the memory required to store the wavefunction. In graphene-loop-mesh.py, the spacing between the b-spline points is varied uniformly. The mesh spacing is a prefactor to the linear spacing between the spline points, so the memory usage goes as the cube of the meshfactor. When you run the calculations, examine the .s000.scalar.dat files with qmca to determine the lowest possible mesh spacing that preserves both the vmc energy and the variance. Similarly, the script graphene-loop-buffer.py uses a feature which generates two spline tables for the wavefunction. One will have half of the mesh spacing requested in the input file and will be valid everywhere. The second one will only be defined in the smallest parallelpiped that contains all of the atoms in the simulation cell with minimum distance given by the buffer size. Again, see what the smallest possible buffer size is that preserves the vmc energy and variance.
Finally, edit the file graphene-final.py which will perform two DMC calculations. In the first, (qmc1) replace the following lines:
\begin{lstlisting}
\begin{shaded}
\begin{verbatim}
meshfactor = xxx,
precision = '---',
truncate = False,
buffer = 0.0,
\end{lstlisting}
\end{verbatim}
\end{shaded}
using the values you have determined to perform the calculation with as small as possible of wavefunction. Note that we can also use single precision arithmetic to store the wavefunction by specifying precision='single'. When you run the script, compare the output of the two DMC calculations in terms of energy and variance. Also see if you can calculate the fraction of memory that you were able to save by using a meshfactor other than 1, a buffer table and single precision arithmetic.
\section{Conclusion}
Upon completion of this lab, you should be able to use the project suite to perform DMC calculations on periodic solids when provided with a pseudopotential. You should also be able to reduce the size of the wavefunction in a solid state calculation in cases where memory is a limiting factor.
Upon completion of this lab, you should be able to use Nexus to perform DMC calculations on periodic solids when provided with a pseudopotential. You should also be able to reduce the size of the wavefunction in a solid state calculation in cases where memory is a limiting factor.
\section{Acknowledgment}
This tutorial was created with support from Sandia National Laboratories.