HOW TO

For the creation of the data set and the evaluation our local server was used and for training we used
a HPC cluster. For details we refer to appendix A.3. We used Ubuntu 20.04 as operation system
with Paraview version 9.10 (Ahrens et al., 2005), Python 3.8.10, PyTorch 1.8.1 (Paszke et al., 2019),
CUDA 11.1 (NVIDIA et al., 2020) and Elmer FEM 8.4 (elm, 2021). In the supplementary materials
you can find code to reproduce our results. First you have to create the data set and afterwards you
can run the training and evaluation script to get the desired results.

CREATION OF DATA SET

To create the necessary data. We have to solve the underlying numeric problem with Elmer FEM.
This is done running the Python-script:

python3 data_set/startElmer.py

In the file “data set/case sif” you can find the necessary parameters for the Elmer FEM. After run-
ning Elmer FEM you will find a folder named “generated data” containing the solutions of every
simulation of every time step in a Paraview compatible file format (.vtu). To note at this point no
human readable images are produced. To change this you have to run

pvpython data_set/startParaview.py

with the Python-environment of Paraview called “pvpython”. After this the data is places in variant
sub folders. The whole “generated data” folder can now be transferred into “cGAN/data” in the
supplementary materials.

TRAINING AND EVALUATION

To start the training with CUDA with subsequent evaluation run

python3 train.py --cuda --evaluate

In “cGAN/config.py” you can find the settings of the script. Here you can decide whether you can
run the cGAN with LSTM or without. You can also specify the number of feature maps of the
generator. The output is produced to the folder “cGAN/model output”.
