Ogoke, F., Meidani, K., Hashemi, A., & Farimani, A. B. (2021). Graph convolutional networks applied to unstructured flow field data. Machine Learning: Science and Technology, 2(4), 045020.
This repository is the implementation corresponding to the paper "Graph convolutional networks applied to unstructured flow field data", linked here.
This project applies Graph Convolutional Neural Networks as a framework for dealing with fluid field data on unstructured grids, enabling inference where traditional image-based analysis methods (e.g. traditional Convolutional Neural Networks) are less applicable.
The data associated with this paper was created through laminar flow field simulations of the aerodynamics of 1200 airfoils, obtained from the UIUC Airfoil Coordinate Database (link). These simulations were carried out using the DOLFIN PDE solver, distributed by FEniCS. This software is presented in the following paper:
- A. Logg and G. N. Wells. DOLFIN: Automated Finite Element Computing, ACM Transactions on Mathematical Software 37 (2010) ACM.
The resulting velocity field and edge connectivity are extracted from the computational mesh for each airfoil. The complete, post-processed dataset of edge connectivity and velocity information can be downloaded directly here These data files contain the edge connectivty and velocity information for the zero angle of attack (AoA) dataset in the folders edges/ and velocity/ respectively, while the multiple angle of attack (AoA) dataset are listed as edges_large/ and velocity_large. The samples are denoted based on their filename from the UIUC database. Additionally, samples that have been post processed for use with traditional ML methods are also included. The [n]manadjvel/ series contains data where the adjacency matrix is multiplied by the velocity node information, and the closest n nodes to the center of the airfoil (based on Manhattan distance) are used for the prediction task. The [n][...]pool series contains edge and velocity information for subset of n nodes selected through random sampling.
The following packages are required in order to run the associated code:
torch_geometric==2.0.3keras==2.6.0tqdm==4.62.3torch==1.8.1+cu101tensorflow==2.8.0
These packages can be installed independently, or all at once by running pip install -r requirements.txt. We recommend that these packages are installed in a new conda environment to avoid clashes with existing package installations. Instructions on defining a new conda environment can be found , and more information on the Pytorch-Geometric installation process can also be found
.
The GCNN is contained in AirfoilGCNN/airfoilgcnn.py, and can be run directly with python airfoilgcnn.py. Upon running the python file, the user will be prompted for the dataset of choice for training, as well as the pre-processing options before the training begins. airfoilmlp.py, airfoilcnn.py, and airfoilshallowmethods.py contain code for the same prediction task, using fully connected neural networks, a convolutional neural network, and a suite of shallow machine learning methods for comparison. The predictions of the GCNN model are saved in the results/ folder for further analysis.