EngineBench

Project page | Link to paper

This is the code relating to the EngineBench database, a collection of datasets collated by the Oxford TPSRG specially for machine learning research in thermal propulsion systems. EngineBench is comprised of Particle Image Velocimetry (PIV) data from different experiments previously run on the transparent combustion chamber (TCC-III) optical engine by General Motors and the University of Michigan.

Table of Contents

1. Acknowledgements
2. Quickstart
3. Installation
4. Usage

Acknowledgements

Publications arising from the use of EngineBench should cite:

1. The following publication:

   @misc{baker2024enginebench,
      title={EngineBench: Flow Reconstruction in the Transparent Combustion Chamber III Optical Engine}, 
      author={Samuel J. Baker and Michael A. Hobley and Isabel Scherl and Xiaohang Fang and Felix C. P. Leach and Martin H. Davy},
      year={2024},
      eprint={2406.03325},
      archivePrefix={arXiv},
      primaryClass={physics.flu-dyn}
   }

2. The EngineBench DOI:

   @misc{samuel_baker_michael_hobley_isabel_scherl_xiaohang_fang_felix_leach_martin_davy_2024,
      title={EngineBench},
      url={https://www.kaggle.com/ds/5000332},
      DOI={10.34740/KAGGLE/DS/5000332},
      publisher={Kaggle},
      author={Samuel Baker and Michael Hobley and Isabel Scherl and Xiaohang Fang and Felix Leach and Martin Davy},
      year={2024}
   }

3. Include the following acknowledgment to the original data source:

   "The TCC engine work has been funded by General Motors through the General Motors University of Michigan Automotive Cooperative Research Laboratory, Engine Systems Division."

This work uses neural network implementations from other sources:

• UNet and UNETR: Project MONAI
• Context encoder GAN: BoyuanJiang (model.py file archived to inpainting/external/ 20-05-2024)
• Convolutional neural operator (CNO): camlab-ethz (CNO2d.py file archived to inpainting/external/ 16-08-2024)

Quickstart

The quickest way to start using EngineBench is via our tutorials using Kaggle notebooks:

1. Browse the data
2. Test different gap types for inpainting
3. Train an inpainting model

Installation

Requirements

• Python 3.8+
• See environment.yml for a full list of dependencies.

Install

git clone https://github.com/...
cd EngineBench
conda env create -f environment.yml

Usage

Data Download

Download the dataset from one of the Kaggle repositories:

• Small dataset
• Full dataset

Training Configuration

Choose the training configurations by editing or creating a new .yaml file in the inpainting/configs/ directory.

Model Training

To train a model, use one of the following scripts based on the model type defined in your config file: For a UNet, UNETR or CNO model:

python train.py --config configs/test_config.yaml

For a GAN model:

python train_gan.py --config configs/test_config.yaml

Model evaluation

Evaluate either model using:

python test.py --config configs/test_config.yaml