Machine learning for rapid discovery of laminar flow channel wall modifications that enhance heat transfer

Installation

Clone this repository and install it in editable mode into a conda environment of your choice using:

pip install -e ChemEngML/

Usage

To run the file:

python3 train.py train_config.yaml --testing True

Shap analysis

Regarding SHAP library: https://github.com/shap/shap https://shap-lrjball.readthedocs.io/en/latest/generated/shap.DeepExplainer.html#shap.DeepExplainer.explain_row

Regarding the codes: main.py: contains the main code and you can run it by typing "python3 main.py"

Two things you have to manuallly modify to run the code:

1. Input the "model_base_path" in mlModel.py: location of the ML model
2. Input the location of the "X_train.npy" in hM_shap.py: location of the dataset used for training the ML model

ChemEngSim (simulations)

When ChemEngSim is located in ./:

• Create a folder for the channel structures e.g. with mkdir channels_1
• Create a folder for logging e.g. with mkdir log
• Copy the config.yml to your submit directory and set your parameters, especially the relative paths to the above created directories, as well as to the ChemEngSim package folder. (channels_i, log, ChemEngSim)

The general workflow is structured with the scripts in ./ChemEngSim/bash_scripts.

1. Create 2D channel geometries

./ChemEngSim/bash_scripts/create_channels.sh config.yml

This will write folder with increasing integers to the base_dir defined in th config. Each folder contains the 2D-channel geometry as hdf5 file and as binary for the fortran code (ibm.bin), as well as a png of the structure.

2. Start a watcher job for early stopping of the calculations

./ChemEngSim/bash_scripts/observe_folder.sh config.yml

This will observe the base_dir and check whether a calculation is running on any of those structures. If this is the case, it will check the history.out file regularly and stop the calculation if the defined level of convergence is reached by writing a stop.now file to the respective directory.

3. Directly after starting the watcher, submit individual job batches

./ChemEngSim/bash_scripts/submit_simson_batch.sh config.yml <n> <step>

The first number defines the number of the simulation batch and the second number the step size. In turn 0 100 would calculate channels 0 to 99, 1 100 would calculate channels 100-199, and so on. Depending on the number of generated channels you have to submit e.g. 30 of those jobs. Make sure that they can finish within 71h and 30min given your defined time limit in the config. E.g. if it is 30 min. your step size should be 143 (71*2+1).
If you can not submit enough jobs to calculate all structures at once, simply start a watcher again after the other jobs finished and submit the remaining jobs. Before you do so, it might be nice to copy the log file of the first watcher from ./log/early_stopping_channels_1.csv to a different location or simply rename it so it is not overwritten. In that way all calculation times are still stored and you can merge the files later on.

4. Collect the output data

./ChemEngSim/bash_scripts/postprocess.sh config.yml

This will collect all 2D channel geometries into one features.h5 and calculate Stanton numbers (St) and drag coefficients (Cf) into results.h5 and results.csv, both located in your base_dir.