README

Replication of results presented in: "Adjoint optimization of pressurized membrane structures using automatic differentiation tools"

REPLICATION OF RESULTS

You must have dolfin-adjoint with FEniCS and IPOPT installed: The easiest way is to use Docker (or Singularity on a research cluster) See http://www.dolfin-adjoint.org/en/latest/download/index.html for detailed instructions

and also https://fenics.readthedocs.io/projects/containers/en/latest/introduction.html

To generate the plots, you will need to have installed matplotlib and pandas

Some visualizations were made externally using Paraview www.paraview.org

Quickstart

You will need Docker installed on your machine to continue.

First, clone the repo: git clone https://github.com/oaniewiarowski/FenicsMembranesCMAME.git

cd FenicsMembranesCMAME

Then start a Docker session with access to current folder:

docker run -ti -v $(pwd):/home/fenics/shared quay.io/dolfinadjoint/pyadjoint

Enter the repo:

cd shared

Install the membrane library (use -e (editable) flag if you want to make any changes):

pip3 install --user -e .

We also need these dependencies: pip3 install --user matplotlib pip3 install --user pandas

The examples are in the following directory:

cd CMAME/examples

Run a problem. The following command runs problem 1, the pressure minimization problem:

python3 main_results.py p1

then plot the results:

python3 main_results.py p1 --plot

The plots are saved in submission/. FYI: The first time runnning the code will take significantly longer becuase of one-time compilation.

Special instructions for Singularity

The code can also be run on a research cluster with resource management ie SLURM using Singularity. To convert the docker image to singularity:

singularity pull docker://quay.io/dolfinadjoint/pyadjoint:2019.1.0

Then you can start a singularity shell: singularity shell pyadjoint_2019.1.0.sif

Singularity automatically exposes the host environment (unlike docker). Next, while in the singularity shell, install the dependencies as indicated above. Singularity> pip3 install --user -e .

To run batch jobs, exit the container (exit), and run the python script main_results.py (the autogenerated slurm script will launch the container) :

python3 main_results.py p1 --slurm

Note: you must configure the slurm script generator settings at the top of main_results.py with your email, time required etc.

To run all the examples at once: python3 main_results.py all --slurm

To plot, launch the singularity shell again. singularity shell ../../../pyadjoint_2019.1.0.sif Singularity> python3 main_results.py p1 --plot

Details

The problem data from the paper are configured in the convenience script: main_results.py.

The script is called from the command line with a command corresponding to the problem ie python3 main_results.py p1

The above command will recreate the results for the plane strain pressure minimization problem (Problem 1).

If running on a cluster, this script will automatically create and submit slurm scripts. If so, the path_to_sif variable in 'main_results.py' may need to be edited to provide the correct path to the .sif file

If running on a personal machine, stdout is redirected to the output file specified by the --path variable

Once the optimization terminates, the results can be plotted:

python3 main_results.py p1 --plot

Problems 1 and 2A have the option to also calculate and save the gradient for the chosen control bounds.

python3 main_results.py p1 --deriv

To run the Taylor tests: python3 main_results.py p1 --tt

To disable optimization (if you only want to run the taylor test and/or gradient study):

python3 main_results.py p1 --tt --deriv --no-optimize python3 main_results.py p1 --tt --no-optimize

See --help for more choices.

For convenience the main results can be replicated with one command: python3 main_results.py all

To experiment with the code, you can also pass other args to the individual problem scripts directly: python3 problem1.py [<options>]

See each problem for details.

Note on runtimes

As currently configured, problem P1 (pressure optimization) and problem 3A (plane strain thickness optimization) each take ~<10 min to run. Each shape optimization (P2 A and B) take about 30 min each. 3D thickness optimization takes about an hour (as it is the largest problem). You can reduce the mesh resolution in the main_results.py to reduce the runtimes, but results may vary/are not guaranteed. You can inspect the stdout (redirected to file specified in path) to monitor the progress.