Dataset Card for PoreSimNet-Data

This is the HuggingFace storage location for all of the training data for the PoreSimNet project.

The PoreSimNet GitHub repository can be found here.

The PoreSimNet Models repository can be found here.

Dataset Details

Dataset Description

This dataset contains both 2D and 3D simulations from running computational fluid dynamics (CFD) simulations through porous media. All of the simulations were written in FLATiron FLOWLab's finite-element CFD library built on top of FEniCS.

It contains roughly 5,000 2D simulations of 2mm x 1mm porous media flows simulating passive contrast agent as it passes through the interstitial space of each simulation.

• Curated by: Josh Gregory

Uses

These data are intended to be used in the biomedical flow and broader porous flows community to help predict permeabilities from porous simulations, such as those found in blood clots or other porous media.

Direct Use

These data are intended to be used in environments with low Reynolds numbers (Re < 10).

Out-of-Scope Use

As of this publication, these models are not intended to treat, diagnose, or prevent any disease and have not been approved by the FDA.

Dataset Structure

MLPs

In 2D, the following applicable MLP training data files are:

• mlp_data_5k (both CSV and Parquet versions): Contain first-order and higher-order texture image features for 5,000 2D porous media flows.
• mlp_data_test (both CSV and Parquet versions): Contain the full evaluation dataset (~1000 simulations) to evaluate the trained MLPs on.

There are also two special files, with both CSV and Parquet versions:

• mlp_data_aug_vert_horiz_flip: This contains the evaluation dataset that has undergone image augmentation before image feature extraction has been performed. Specifically, vertical and horizontal flips with a 50% probability of each.
• mlp_data_aug_cnn_pipe: This contains the evaluation dataset that has undergone image augmentation before image feature extraction has been performed in the same pipeline as the CNN networks.

In 3D, the following applicable MLP training data files are:

• mlp_data_3d_100 (both CSV and Parquet versions): Contain first-order and higher-order texture image features for 100 3D porous media flows.
• mlp_data_3d_100 (both CSV and Parquet versions): Contain first-order and higher-order texture image features for 300 3D porous media flows.

CNNs

Each CNN data source is in two file formats:

• .lmdb: Stores the images in Lightning Memory-Mapped Database (LMDB) format for efficient network processing.
• .tar.gz: Regular storage of each raw image, with corresponding permeabilities embedded in each image's EXIF metadata as a user commment.

The directories ending in only _crop indicate that the image is rectangular (i.e., only white space has been removed). File names that end in center_crop indicate that a 512x512 pixel center crop has been taken from the original image (the original image being contained in the directories that end in _crop).

The directories that contain the word test in them correspond to the evaluation dataset and were used to benchmark the CNN models.

No 3D data was used for the CNNs.

Annotation process

Each simulation had its permeability calcualted using Darcy's Law. For the MLP networks, this became a column labeled k, and for each CNN network was embedded into each image's EXIF metadata as a user comment.

Personal and Sensitive Information

This dataset does not include any personal and sensitive information.

Bias, Risks, and Limitations

Each simulation has flow entering with a fully-developed profile with a peak inlet velocity of 20 mm/s. While based on literature values for blood flow, it is recommended that ranges of values and angles of this blood flow be explored.

Dataset Card Authors

Josh Gregory: josh.a.gregory42@gmail.com

Dataset Card Contact

Debanjan Mukherjee: debanjan@colorado.edu