📄 Optimizing progress variables for ammonia/hydrogen combustion using encoding-decoding networks

This repository contains code, datasets, and results from the paper:

Kamila Zdybał, James C. Sutherland, Alessandro Parente - Optimizing progress variables for ammonia/hydrogen combustion using encoding-decoding networks, 2024.

Data and results files

Data and results files will be shared separately via GoogleDrive as they take over 5GB of space.

Script for loading data ammonia-Stagni-load-data.py

Requirements

We have used Python==3.10.13 and the following versions of all libraries:

pip install numpy==1.26.2
pip install pandas==2.1.3
pip install scipy==1.11.4
pip install scikit-learn==1.3.2
pip install tensorflow==2.15.0
pip install keras==2.15.0

You will also need our library PCAfold==2.2.0.

Other requirements are:

pip install matplotlib
pip install plotly
pip install cmcrameri

Python scripts

The order of executing scripts

First, run the PV optimization with RUN-PV-optimization.py with desired parameters. Once you have the results files, you can run quantitative assessment of PVs with RUN-VarianceData.py. Both those scripts load the appropriate data under the hood using ammonia-Stagni-load-data.py.

You have a lot of flexibility in setting different ANN hyper-parameters in those two scripts using the argparse Python library. If you're new to argparse, check out my short video tutorials:

Optimizing PVs

Master script for running PV optimization RUN-PV-optimization.py

The above script uses one of the following under the hood:

QoI-aware encoder-decoder for the $(f, PV)$ optimization QoI-aware-ED-f-PV.py
QoI-aware encoder-decoder for the $(f, PV, \gamma)$ optimization QoI-aware-ED-f-PV-h.py

depending on which --parameterization you selected.

Quantitative assessment of PVs

Master script for running PV optimization RUN-VarianceData.py

The above script uses one of the following under the hood:

Assessment of $(f, PV)$ parameterizations VarianceData-f-PV.py
Assessment of $(f, PV, \gamma)$ parameterizations VarianceData-f-PV-h.py

depending on which --parameterization you selected.

Running Python jobs

This is a minimal example for running a Python script with all hyper-parameters set as per §2.2 in the paper:

python RUN-PV-optimization.py --parameterization 'f-PV' --data_type 'SLF' --data_tag 'NH3-H2-air-25perc' --random_seeds_tuple 0 20 --target_variables_indices 0 1 3 5 6 9

Alternatively, you can change various parameters (kernel initializer, learning rate, etc.) using the appropriate argument:

python RUN-PV-optimization.py --initializer 'GlorotUniform' --init_lr 0.001 --parameterization 'f-PV' --data_type 'SLF' --data_tag 'NH3-H2-air-25perc' --random_seeds_tuple 0 20 --target_variables_indices 0 1 3 5 6 9

If you'd like to remove pure stream components from the PV definition (non-trainable pure streams preprocessing as discussed in §3.4. in the paper) use the flag:

--no-pure_streams

as an extra argument.

To run $(f, PV)$ optimization, use:

--parameterization 'f-PV'

To run $(f, PV, \gamma)$ optimization, use:

--parameterization 'f-PV-h'

Note: Logging with Weights & Biases is also possible in the scripts above.

Jupyter notebooks

All results are post-processed and visualized in dedicated Jupyter notebooks. You can access the appropriate notebook below:

Reproducing Figs. 2-3 - Quantitative assessment of the optimized PVs

→ This Jupyter notebook can be used to reproduce Figs. 2-3.

Reproducing Fig. 4 and Fig. 10 - Physical insight into the optimized PVs

→ This Jupyter notebook can be used to reproduce Fig. 4 and Fig. 10.

Reproducing supplementary Figs. S37-S38 - The effect of scaling encoder inputs prior to training

→ This Jupyter notebook can be used to reproduce supplementary Figs. S37-S38.

kamilazdybal/pv-optimization