`magudi` tutorial

Prerequisite

Fortran compiler: ifort or gfortran
MPI compiler: mvapich or openmpi

MPI compiler must be installed with the Fortran compiler that compiles magudi.

Installation

Say the source code directory (mostly magudi) is named as <magudi-src>. We recommend making a separate directory for the installation, which we denote <magudi-build>. The installation is then executed via the following commands on the terminal,

mkdir <magudi-build>
cd <magudi-build>
cmake -D<cmake-var>=% <magudi-src>
make

The flag -D<cmake-var>=% is optional, to manually set <cmake-var> to %. Some useful flags <cmake-var> are:

CMAKE_BUILD_TYPE: set to either release or debug
CMAKE_Fortran_COMPILER: set to a specific fortran compiler binary file

Example: acoustic monopole

Configuration

We copy the example directory <magudi-src>/examples/AcousticMonopole to any location you like.

cp -r <magudi-src>/examples/AcousticMonopole/* ./

config.py file includes all routines to generate grid, function, and flow solution files in PLOT3D format. This needs a python file from <magudi-src>/utils/python.

cp <magudi-src>/utils/python/plot3dnasa.py ./

Note that plot3dnasa.py and config.py are written in python2. They are currently not compatible with python3. You can either import config.py and use the routines, or run config.py itself:

python2 config.py

This will generate:

Grid file: AcousticMonopole.xyz
Initial condition file: AcousticMonopole.ic.q
Function files: AcousticMonopole.target_mollifier.f, AcousticMonopole.control_mollifier.f

Baseline simulation

After compiling, main binary executables are generated in <magudi-build>/bin/. There are also auxiliary executables in <magudi-build>/utils/. Check the codes to see what tasks they do.

For baseline simulation, make a link to <magudi-build>/bin/forward:

ln -s <magudi-build>/bin/forward ./

forward requires two input files:

magudi.inp: this includes all flags/parameters for simulation.
bc.dat: this includes boundary condition information on grids. you can specify your own boundary condition file in magudi.inp.

To execute the baseline simulation, run:

./forward --output J0.txt

This saves the quantity-of-interest (QoI) J in J0.txt. Additionally, solutions at designated timesteps are saved. These will be required for adjoint simulation. Please check the code to see more optional arguments for forward.

Adjoint simulation

For baseline simulation, make a link to <magudi-build>/bin/adjoint:

ln -s <magudi-build>/bin/adjoint ./

To execute the adjoint simulation, run:

./adjoint --output Grad0.txt

This saves the gradient magnitude in Grad0.txt. Gradient vector (forcing) will be saved as AcousticMonopole.gradient.controlRegion.dat. Additionally, solutions at designated timesteps are saved. These will be required for adjoint simulation. Please check the code to see more optional arguments for adjoint.

Gradient accuracy check by finite-difference

We check the gradient accuracy by applying the control forcing along the gradient direction. In magudi.inp, change the controller_switch flag,

controller_switch = true

To make the control forcing file, make a link to <magudi-build>/bin/zaxpy:

ln -s <magudi-build>/bin/zaxpy ./

zaxpy refers to z=a*x+y, where z, x, and y are the vector .dat files with the same length of the gradient file, and a is a scalar. We make a control forcing that is 0.0001 times smaller than the gradient,

./zaxpy AcousticMonopole.control_forcing.controlRegion.dat 1e-4 AcousticMonopole.gradient.controlRegion.dat`

where the argument y is automatically taken to be zero. Run another forward run with this file,

./forward --output J1.txt

where QoI is saved in J1.txt. A python3 routine to compute the finite-difference and check the accuracy is

import numpy as np
fID = open('J0.txt','r')
J0 = float(fID.read())
fID.close()
fID = open('Grad0.txt','r')
Grad0 = float(fID.read())
fID.close()
fID = open('J1.txt','r')
J1 = float(fID.read())
fID.close()

A1 = 1.0E-4
Grad1 = (J1-J0) / A1
Error = abs( (Grad1 - Grad0)/Grad0 )
print ("{:.16E}".format(A1), "{:.16E}".format(Error))

Following python3 subroutine does this job for multiple amplitudes and save the errors in AcousticMonopole.gradient_accuracy.txt:

import numpy as np
import subprocess

fID = open('J0.txt','r')
QoI0 = float(fID.read())
fID.close()
fID = open('Grad0.txt','r')
Grad0 = float(fID.read())
fID.close()

Nk = 20
Ak = 10.0**(-2.0-0.25*np.array(range(Nk)))
QoIk = np.zeros((Nk,),dtype=np.double)
Gradk = np.zeros((Nk,),dtype=np.double)
ek = np.zeros((Nk,),dtype=np.double)

for k in range(Nk):
    amp = Ak[k]
    command = ''
    command += './zaxpy AcousticMonopole.control_forcing.controlRegion.dat %.16E AcousticMonopole.gradient.controlRegion.dat`' % amp
    command += './forward --output J1.txt'
    fID = open('test-step.sh','w')
    fID.write(command)
    fID.close()
    subprocess.check_call('bash test-step.sh', shell=True)

    fID = open('J1.txt','r')
    QoIk[k] = float(fID.read())
    fID.close()

    Gradk[k] = (QoIk[k]-QoI0)/Ak[k]
    ek[k] = abs( (Gradk[k]-Grad0)/Grad0 )
    print ("{:.16E}".format(Ak[k]), "{:.16E}".format(QoIk[k]), "{:.16E}".format(Gradk[k]), "{:.16E}".format(ek[k]))

    fId = open(globalPrefix+'.gradient_accuracy.txt','a+')
    fId.write('%.16E\t%.16E\t%.16E\t%.16E\n' % (Ak[k], QoIk[k], Gradk[k], ek[k]))
    fId.close()

dreamer2368/magudi

magudi tutorial