MegaFlow2D 0.3.5

MegaFlow2D: A Large-Scale Dataset for 2D Flow Simulation

MegaFlow2D

Overview

The MegaFlow2D dataset package of parameteric CFD simulation results for machine learning / super-resolution purposes.

The package contains:

1. A standard structure for transferring simulation results into graph structure.
2. Common utility functions for visualizing, retrieving and processing simulation results. (Everything that requires the FEniCS or dolfin package can only be run on linux or wsl.)

Installation

The MegaFlow dataset can be installed by pip:

pip install MegaFlow2D

Running pip install would automatically configure package dependencies, however to build graphical models torch-geometric needs to be installed manually.

Using the MegaFlow package

The MegaFlow package provides a simple interface for initializing and loading the dataset.

from megaflow.dataset import MegaFlow2D

if __name__ == '__main__':
    dataset = MegaFlow2D(root='/path/to/your/directory', download=True)
    # if the dataset is not processed, the process function will be called automatically. 
    # to facilitate multi-thread processing, be sure to exceute the process function in '__main__'.

    # get one sample
    sample = dataset.get(0)
    print('Number of nodes: {}, number of edges: {}'.format(sample.num_nodes, sample.num_edges))

Using the example scripts

We provide an example script for training a super-resolution model on the MegaFlow2D dataset. The script can be found in the examples directory. The script can be run by (one configuration example):

python examples/train.py --root /path/to/your/directory --dataset MegaFlow2D --tranform normalize --model FlowMLError --epochs 100 --batch_size 32