How to Converge the CUTOFF and REL_CUTOFF
Introduction
QUICKSTEP, as with nearly all ab initio Density Functional Theory simulation packages, requires
the use of a real-space (RS) integration grid to represent certain functions, such as the electron
density and the product Gaussian functions. QUICKSTEP uses a multi-grid system for mapping the
product Gaussians onto the RS grid(s), so that wide and smooth Gaussian functions are mapped onto a
coarser grid than narrow and sharp Gaussians. The electron density is always mapped onto the finest
grid.
Choosing a fine enough integration grid for a calculation is crucial in obtaining meaningful and accurate results. In this tutorial, we will show the reader how to systematically find the correct settings for obtaining a sufficiently fine integration grid for his/her calculation.
This tutorial assumes the reader already has some knowledge of how to perform a simple energy
calculation using QUICKSTEP (this can be found in tutorial:
Calculating Energy and Forces using Quickstep).
A completed example from an earlier calculation can be obtained from here. The calculations were carried out using CP2K version 2.4.
“QUICKSTEP” Multi-Grid
Before we go through the input file, it is worthwhile to explain how the multi-grid is constructed in QUICKSTEP, and how the Gaussians are mapped onto the different grid levels. Hopefully this will offer the reader a clear picture of how the key control parameters affect the grids, and thus the overall accuracy of a calculation.
All multi-grid related settings for a calculation is controlled via keywords in MULTIGRID subsection of DFT subsection in FORCE_EVAL. The number of levels for the multi-grid is defined by NGRIDS, and by default this is set to 4. The keyword CUTOFF defines the planewave cutoff (default unit is in Ry) for the finest level of the multi-grid. The higher the planewave cutoff, the finer the grid. The corresponding planewave cutoffs for the subsequent grid levels (from finer to coarser) are defined by the formula:
where \(\alpha\) has a default value of 3.0, and since CP2K versions 2.0, can be configured by the
keyword PROGRESSION_FACTOR(#CP2K_INPUT.FORCE_EVAL.DFT.MGRID.PROGRESSION_FACTOR). Therefore, the
higher the value of CUTOFF the finer grid for all
multi-grid levels.
Having constructed the multi-grid, QUICKSTEP then needs to map the Gaussians onto the grids. The
keyword REL_CUTOFF controls which product Gaussians
are mapped onto which level of the multi-grid. CP2K tries to map each Gaussian onto a grid such
that the number of grid points covered by the Gaussian—no matter how wide or narrow—are roughly the
same. REL_CUTOFF defines the planewave cutoff of a
reference grid covered by a Gaussian with unit standard deviation
\((e^{\mathrm{|\overrightarrow{r}|^{2}}})\). A Gaussian is mapped onto the coarsest level of the
multi-grid, on which the function will cover number of grid points greater than or equal to the
number of grid points \(e^{\mathrm{|\overrightarrow{r}|^{2}}}\) will cover on a reference grid defined
by REL_CUTOFF.
Therefore, the two most important keywords effecting the integration grid and the accuracy of a calculation are CUTOFF and REL_CUTOFF. If CUTOFF is too low, then all grids will be coarse and the calculation may become inaccurate; and if REL_CUTOFF is too low, then even if you have a high CUTOFF, all Gaussians will be mapped onto the coarsest level of the multi-grid, and thus the effective integration grid for the calculation may still be too coarse.
Example: Bulk Si with 8 atoms in a cubic cell
We demonstrate the process using an example based on Bulk Si with 8 atoms in a face centred cubic unit cell.
Template Input File
To systematically find the best CUTOFF and REL_CUTOFF values which are sufficient for a given accuracy (say, \(10^{−6}\) Ry in total energy), we need to perform a series of single point energy calculations. It is much easier to use a set of scripts that can automate this process.
To do this, we first write a template input file: template.inp, as shown below:
&GLOBAL
PROJECT Si_bulk8
RUN_TYPE ENERGY
PRINT_LEVEL MEDIUM
&END GLOBAL
&FORCE_EVAL
METHOD Quickstep
&DFT
BASIS_SET_FILE_NAME BASIS_SET
POTENTIAL_FILE_NAME GTH_POTENTIALS
&MGRID
NGRIDS 4
CUTOFF LT_cutoff
REL_CUTOFF LT_rel_cutoff
&END MGRID
&QS
EPS_DEFAULT 1.0E-10
&END QS
&SCF
SCF_GUESS ATOMIC
EPS_SCF 1.0E-6
MAX_SCF 1
ADDED_MOS 10
CHOLESKY INVERSE
IGNORE_CONVERGENCE_FAILURE TRUE
&SMEAR ON
METHOD FERMI_DIRAC
ELECTRONIC_TEMPERATURE [K] 300
&END SMEAR
&DIAGONALIZATION
ALGORITHM STANDARD
&END DIAGONALIZATION
&MIXING
METHOD BROYDEN_MIXING
ALPHA 0.4
BETA 0.5
NBROYDEN 8
&END MIXING
&END SCF
&XC
&XC_FUNCTIONAL PADE
&END XC_FUNCTIONAL
&END XC
&END DFT
&SUBSYS
&KIND Si
ELEMENT Si
BASIS_SET SZV-GTH-PADE
POTENTIAL GTH-PADE-q4
&END KIND
&CELL
SYMMETRY CUBIC
A 5.430697500 0.000000000 0.000000000
B 0.000000000 5.430697500 0.000000000
C 0.000000000 0.000000000 5.430697500
&END CELL
&COORD
Si 0.000000000 0.000000000 0.000000000
Si 0.000000000 2.715348700 2.715348700
Si 2.715348700 2.715348700 0.000000000
Si 2.715348700 0.000000000 2.715348700
Si 4.073023100 1.357674400 4.073023100
Si 1.357674400 1.357674400 1.357674400
Si 1.357674400 4.073023100 4.073023100
Si 4.073023100 4.073023100 1.357674400
&END COORD
&END SUBSYS
&PRINT
&TOTAL_NUMBERS ON
&END TOTAL_NUMBERS
&END PRINT
&END FORCE_EVAL
We go through this input file quickly. Readers who have gone through the
tutorial on how to perform a simple static energy and force calculation
using QUICKSTEP should have no trouble in understanding most parts the above input.
Some noticeable settings are:
&GLOBAL
PROJECT Si_bulk8
RUN_TYPE ENERGY
PRINT_LEVEL MEDIUM
&END GLOBAL
The keyword RUN_TYPE is set to ENERGY, this tells CP2K to only
calculate the energies of the system, forces will not be calculated. Since we are only interested in
the convergence of the integration grid, just looking at the total energy usually suffices; and
since we will be performing a series of computations, the cheaper each run is the better. We set
PRINT_LEVEL to MEDIUM, so that the information about how many
Gaussian functions are mapped onto which grid are printed. We need this information to analyse the
suitability of the chosen REL_CUTOFF value.
The most important part in the template input is:
&MGRID
NGRIDS 4
CUTOFF LT_cutoff
REL_CUTOFF LT_rel_cutoff
&END MGRID
The symbols LT_cutoff and LT_rel_cutoff are markers, which the automated scripts will search for
and replace with the relevant values. The default units for both
CUTOFF and
REL_CUTOFF are Ry.
In SCF subsection, we have set
MAX_SCF 1
So that no self-consistent loops will be performed. This is okay for checking the integration grid, because irrespective of self-consistency, grid settings with fine enough meshes should give consistent energies.
Converging ‘‘CUTOFF’’
We start by setting REL_CUTOFF to a relatively high number, and systematically vary CUTOFF. Setting REL_CUTOFF to 60 Ry is usually sufficient for most calculations, and in any case this will be checked later when we vary REL_CUTOFF.
Generating Inputs
We want to perform a series of calculations, with CUTOFF
ranging from 50 Ry to 500 Ry in steps of 50 Ry. From experience, the desired
CUTOFF for an accuracy of \(10^{−6}\) Ry for the total
energy should be well within this range. To do this, we first need to make sure the basis and
pseudopotential parameter files BASIS_SET and GTH_POTENTIALS (you can copy them from
../cp2k/data/) are in the working directory together with template.inp, then one can write a
bash script, such as the file cutoff_inputs.sh shown below:
#!/bin/bash
cutoffs="50 100 150 200 250 300 350 400 450 500"
basis_file=BASIS_SET
potential_file=GTH_POTENTIALS
template_file=template.inp
input_file=Si_bulk8.inp
rel_cutoff=60
for ii in $cutoffs ; do
work_dir=cutoff_${ii}Ry
if [ ! -d $work_dir ] ; then
mkdir $work_dir
else
rm -r $work_dir/*
fi
sed -e "s/LT_rel_cutoff/${rel_cutoff}/g" \
-e "s/LT_cutoff/${ii}/g" \
$template_file > $work_dir/$input_file
cp $basis_file $work_dir
cp $potential_file $work_dir
done
The user should remember to set the permission of the new script file to be executable:
chmod u+x ./cutoff_inputs.sh
Entering the command line
./cutoff_inputs.sh
generates directories cutoff_50Ry, cutoff_100Ry, …, each containing BASIS_SET,
GTH_POTENTIALS and an input file Si_bulk8.inp, which is exactly the same as template.inp,
except that REL_CUTOFF is set to 60, and
CUTOFF is set to the respective values in the range
between 50 Ry and 500 Ry.
Running Calculations
With the input files generated and checked, the next step is to run them. A bash script such as
cutoff_run.sh shown below does the job:
#!/bin/bash
cutoffs="50 100 150 200 250 300 350 400 450 500"
cp2k_bin=cp2k.popt
input_file=Si_bulk8.inp
output_file=Si_bulk8.out
no_proc_per_calc=2
no_proc_to_use=16
counter=1
max_parallel_calcs=$(expr $no_proc_to_use / $no_proc_per_calc)
for ii in $cutoffs ; do
work_dir=cutoff_${ii}Ry
cd $work_dir
if [ -f $output_file ] ; then
rm $output_file
fi
mpirun -np $no_proc_per_calc --bind-to none $cp2k_bin -o $output_file $input_file &
cd ..
mod_test=$(echo "$counter % $max_parallel_calcs" | bc)
if [ $mod_test -eq 0 ] ; then
wait
fi
counter=$(expr $counter + 1)
done
wait
The above script is slightly complex, because it allows several jobs to run in parallel. Setting the
variable cp2k_bin defines the path to the CP2K binary. In this case, the parallel version
cp2k.popt is found in the system PATH. no_proc_per_calc sets the number of MPI processes to
be used in parallel for each job. no_proc_to_use sets the total number of processors to be used
for running all of the jobs. Since the calculations are launched under one process, the
bind-to none option is required for mpirun to not assign each calculation to the same two
processors. Note that when using a Resource Manager for submitting jobs, e.g. SLURM, this command
should not be used. In the above example, the jobs are run on a 24 core local workstation, a total
of 16 cores are used for performing the CUTOFF
convergence test calculations, and 2 cores are used for each calculation. This means up to 8 jobs
will run in parallel, until the jobs are exhausted from the list given in cutoffs.
The reader can write their own script where they see fit, and if he/she just want the jobs to run in serial, then there is no need for this complexity.
Again
chmod u+x ./cutoff_run.sh
followed by
./cutoff_run.sh &
runs the calculations in the background. This calculation only took a couple of minutes to complete on our local workstation.
Analysing Results
After all of the calculations have finished, all the information about total energies and
distribution of Gaussians on the multi-grid are written in the Si_bulk8.out files in each job
directories.
The total energy can be found in the section of the output shown below (in this example from
cutoff_100Ry/Si_bulk8.out):
SCF WAVEFUNCTION OPTIMIZATION
Step Update method Time Convergence Total energy Change
------------------------------------------------------------------------------
Trace(PS): 32.0000000000
Electronic density on regular grids: -31.9999999980 0.0000000020
Core density on regular grids: 31.9999999944 -0.0000000056
Total charge density on r-space grids: -0.0000000036
Total charge density g-space grids: -0.0000000036
1 NoMix/Diag. 0.40E+00 0.4 1.10090760 -32.3804557631 -3.24E+01
1 NoMix/Diag. 0.40E+00 0.4 1.10090760 -32.3804557631 -3.24E+01
*** SCF run NOT converged ***
Electronic density on regular grids: -31.9999999980 0.0000000020
Core density on regular grids: 31.9999999944 -0.0000000056
Total charge density on r-space grids: -0.0000000036
Total charge density g-space grids: -0.0000000036
Overlap energy of the core charge distribution: 0.00000000005320
Self energy of the core charge distribution: -82.06393942512820
Core Hamiltonian energy: 16.92855916540793
Hartree energy: 42.17635056223367
Exchange-correlation energy: -9.42142606564066
Electronic entropic energy: 0.00000000000000
Fermi energy: 0.00000000000000
Total energy: -32.38045576307407
Regexp search
"^[ \t]*Total energy:"
will find the relevant line.
Similarly, information on distribution of Gaussians on the multi-grid can be found in the section:
-------------------------------------------------------------------------------
---- MULTIGRID INFO ----
-------------------------------------------------------------------------------
count for grid 1: 2720 cutoff [a.u.] 50.00
count for grid 2: 5000 cutoff [a.u.] 16.67
count for grid 3: 2760 cutoff [a.u.] 5.56
count for grid 4: 16 cutoff [a.u.] 1.85
total gridlevel count : 10496
which tells us that for CUTOFF of 100 Ry and
REL_CUTOFF of 60 Ry, 2720 product Gaussians has been
distributed to grid level 1, the finest level, 5000 for level 2, 2760 for level 3 and 16 for level
4, the coarsest. The planewave cutoff for each multi-grid level can be read from the right-hand-side
columns. Here [a.u.] means the Hartree energy unit, 1 Ha = 2 Ry.
It is much easier if we can gather all the information together into one file, which allows us to
plot the results. This can be done, again, by using a simple script. cutoff_analyse.sh shown below
is such an example:
#!/bin/bash
cutoffs="50 100 150 200 250 300 350 400 450 500"
input_file=Si_bulk8.inp
output_file=Si_bulk8.out
plot_file=cutoff_data.ssv
rel_cutoff=60
echo "# Grid cutoff vs total energy" > $plot_file
echo "# Date: $(date)" >> $plot_file
echo "# PWD: $PWD" >> $plot_file
echo "# REL_CUTOFF = $rel_cutoff" >> $plot_file
echo -n "# Cutoff (Ry) | Total Energy (Ha)" >> $plot_file
grid_header=true
for ii in $cutoffs ; do
work_dir=cutoff_${ii}Ry
total_energy=$(grep -e '^[ \t]*Total energy' $work_dir/$output_file | awk '{print $3}')
ngrids=$(grep -e '^[ \t]*QS| Number of grid levels:' $work_dir/$output_file | \
awk '{print $6}')
if $grid_header ; then
for ((igrid=1; igrid <= ngrids; igrid++)) ; do
printf " | NG on grid %d" $igrid >> $plot_file
done
printf "\n" >> $plot_file
grid_header=false
fi
printf "%10.2f %15.10f" $ii $total_energy >> $plot_file
for ((igrid=1; igrid <= ngrids; igrid++)) ; do
grid=$(grep -e '^[ \t]*count for grid' $work_dir/$output_file | \
awk -v igrid=$igrid '(NR == igrid){print $5}')
printf " %6d" $grid >> $plot_file
done
printf "\n" >> $plot_file
done
Type
chmod u+x ./cutoff_analyse.sh
and then run it using
./cutoff_analyse.sh
will produce a file named cutoff_data.ssv, which looks like:
# Grid cutoff vs total energy
# Date: Mon Jan 20 21:20:34 GMT 2014
# PWD: /home/tong/tutorials/converging_grid/sample_output
# REL_CUTOFF = 60
# Cutoff (Ry) | Total Energy (Ha) | NG on grid 1 | NG on grid 2 | NG on grid 3 | NG on grid 4
50.00 -32.3795329864 5048 5432 16 0
100.00 -32.3804557631 2720 5000 2760 16
150.00 -32.3804554850 2032 3016 5432 16
200.00 -32.3804554982 1880 2472 3384 2760
250.00 -32.3804554859 264 4088 3384 2760
300.00 -32.3804554843 264 2456 5000 2776
350.00 -32.3804554846 56 1976 5688 2776
400.00 -32.3804554851 56 1976 3016 5448
450.00 -32.3804554851 0 2032 3016 5448
500.00 -32.3804554850 0 2032 3016 5448
The data shows that given the REL_CUTOFF value of 60 Ry, setting CUTOFF to 250 Ry and above would give an error in total energy less than \(10^{−8}\) Ha. The reader may also notice that as CUTOFF increases, the number of Gaussians being assigned to the finest grids decreases. Therefore, simply increasing CUTOFF without increasing REL_CUTOFF may eventually lead to a slow convergence in energy, as more and more Gaussians get pushed to coarser grid levels, negating the increase in CUTOFF.
In this example, the test results point to 250 Ry as a good choice for CUTOFF, as the total energy is converged, and the distribution of Gaussian functions on the grids are reasonable: it is the lowest cutoff energy where the finest grid level is used, but at the same time with the majority of the Gaussians on the coarser grids.
Converging “REL_CUTOFF”
In the next step, we vary the value of REL_CUTOFF while keeping CUTOFF fixed at 250 Ry.
Generating Inputs
For the energy convergence test with varying
REL_CUTOFF, we follow a similar procedure as that for
CUTOFF. Using the same template input file
template.inp, we can write a script called rel_cutoff_inputs.sh:
#!/bin/bash
rel_cutoffs="10 20 30 40 50 60 70 80 90 100"
basis_file=BASIS_SET
potential_file=GTH_POTENTIALS
template_file=template.inp
input_file=Si_bulk8.inp
cutoff=250
for ii in $rel_cutoffs ; do
work_dir=rel_cutoff_${ii}Ry
if [ ! -d $work_dir ] ; then
mkdir $work_dir
else
rm -r $work_dir/*
fi
sed -e "s/LT_cutoff/${cutoff}/g" \
-e "s/LT_rel_cutoff/${ii}/g" \
$template_file > $work_dir/$input_file
cp $basis_file $work_dir
cp $potential_file $work_dir
done
and again running
chmod u+x ./rel_cutoff_inputs.sh
./rel_cutoff_inputs.sh
Setting the permission for the script to “executable”, and running it produces directories
rel_cutoff_10Ry, rel_cutoff_20Ry, …, each containing files BASIS_SET, GTH_POTENTIALS and an
input Si_bulk8.inp, which is identical to template.inp, except that
CUTOFF is set to 250, and
REL_CUTOFF is set to 10, 20, …, 100 respectively.
Running Calculations
Again to run the calculations, we can use the script rel_cutoff_run.sh, as shown below:
#!/bin/bash
rel_cutoffs="10 20 30 40 50 60 70 80 90 100"
cp2k_bin=cp2k.popt
input_file=Si_bulk8.inp
output_file=Si_bulk8.out
no_proc_per_calc=2
no_proc_to_use=16
counter=1
max_parallel_calcs=$(expr $no_proc_to_use / $no_proc_per_calc)
for ii in $rel_cutoffs ; do
work_dir=rel_cutoff_${ii}Ry
cd $work_dir
if [ -f $output_file ] ; then
rm $output_file
fi
mpirun -np $no_proc_per_calc --bind-to none $cp2k_bin -o $output_file $input_file &
cd ..
mod_test=$(echo "$counter % $max_parallel_calcs" | bc)
if [ $mod_test -eq 0 ] ; then
wait
fi
counter=$(expr $counter + 1)
done
wait
In the above example, again, we have used 16 cores in total, and with each job using 2 MPI
processes. To run the jobs, use:
chmod u+x ./rel_cutoff_run.sh
./rel_cutoff_run.sh &
Total energies and distribution of Gaussian functions on the multi-grid are obtained the same way from the results as that for the CUTOFF calculations.
To put all of the results from the REL_CUTOFF
calculations in one place, we can make some minor modifications to cutoff_analyse.sh and save it
as rel_cutoff_analyse.sh:
#!/bin/bash
rel_cutoffs="10 20 30 40 50 60 70 80 90 100"
input_file=Si_bulk8.inp
output_file=Si_bulk8.out
plot_file=rel_cutoff_data.ssv
cutoff=250
echo "# Rel Grid cutoff vs total energy" > $plot_file
echo "# Date: $(date)" >> $plot_file
echo "# PWD: $PWD" >> $plot_file
echo "# CUTOFF = ${cutoff}" >> $plot_file
echo -n "# Rel Cutoff (Ry) | Total Energy (Ha)" >> $plot_file
grid_header=true
for ii in $rel_cutoffs ; do
work_dir=rel_cutoff_${ii}Ry
total_energy=$(grep -e '^[ \t]*Total energy' $work_dir/$output_file | awk '{print $3}')
ngrids=$(grep -e '^[ \t]*QS| Number of grid levels:' $work_dir/$output_file | \
awk '{print $6}')
if $grid_header ; then
for ((igrid=1; igrid <= ngrids; igrid++)) ; do
printf " | NG on grid %d" $igrid >> $plot_file
done
printf "\n" >> $plot_file
grid_header=false
fi
printf "%10.2f %15.10f" $ii $total_energy >> $plot_file
for ((igrid=1; igrid <= ngrids; igrid++)) ; do
grid=$(grep -e '^[ \t]*count for grid' $work_dir/$output_file | \
awk -v igrid=$igrid '(NR == igrid){print $5}')
printf " %6d" $grid >> $plot_file
done
printf "\n" >> $plot_file
done
Making the script executable, and running the script using
chmod u+x rel_cutoff_analyse.sh
./rel_cutoff_analyse.sh
produces the following results written in file rel_cutoff_data.ssv:
# Rel Grid cutoff vs total energy
# Date: Mon Jan 20 00:45:14 GMT 2014
# PWD: /home/tong/tutorials/converging_grid/sample_output
# CUTOFF = 250
# Rel Cutoff (Ry) | Total Energy (Ha) | NG on grid 1 | NG on grid 2 | NG on grid 3 | NG on grid 4
10.00 -32.3902980020 0 0 2032 8464
20.00 -32.3816384686 0 264 4088 6144
30.00 -32.3805115576 0 2032 3016 5448
40.00 -32.3805116025 56 1976 3016 5448
50.00 -32.3804555002 264 2456 5000 2776
60.00 -32.3804554859 264 4088 3384 2760
70.00 -32.3804554859 1880 2472 3384 2760
80.00 -32.3804554859 1880 2472 3384 2760
90.00 -32.3804554848 2032 3016 5432 16
100.00 -32.3804554848 2032 3016 5432 16
The results show that as one increases the value of REL_CUTOFF, more Gaussians get mapped onto the finer grids. The error in total energy reduces to less than \(10^{−8}\) Ha when REL_CUTOFF is greater or equal to 60 Ry. The results thus indicate that 60 Ry is indeed a suitable choice for the value of REL_CUTOFF.
So finally we conclude that the setting
&MGRID
CUTOFF 250
REL_CUTOFF 60
&END MGRID
is sufficient for a calculation with the required accuracy.