Permeability Prediction in 2D: Dataset and Trained Convolutional Neural Networks

The dataset was generated for the purpose of training convolutional neural network (CNN) models for permeability prediction of 2D structures. The whole dataset is part of a study on predicting permeability using CNNs, while addressing discussions that are largely absent from the current literature, such as the effect of data diversity in the accuracy, input pre-processing, error estimation, architecture comparisons, and sources of error. A link to the publication, which includes a lot more detail about the dataset and CNN models, will be added once it is published.

The data included in this dataset is split into three different folders. The data under the "Training Data" folder includes 4,500 images, divided in 15 sub-folders. Each sub-folder contains 901 files, which are 300 images, a pressure-velocity map for each of the 300 structures, convergence data for each individual structure, and one comma-separated file (csv) summarizing all simulation results in the folder. The pressure and velocity maps together with the convergence information are direct results of the CFD algorithm used, but the important information for training the CNN models are the images and the permeability data in the csv files.

The "Trained CNNs" folder contains all of the trained CNN models as described in the linked publication for predicting permeability. That includes the ensemble of VGG19 networks. The "External Test Set" includes the same type of data as the "Training Data" folder, but this section of data was only used to test the CNN models. In other words, the trained CNN models never saw any of this data in training, only in testing. The "External Test Set" folder also includes data for phase-size distributions and surface area for all data in this repository. For more details on those, refer to the publication.

The CFD code and the image generation code can be found in the following GitHub, along with more extensive documentation:
https://github.com/adama-wzr/PixelBasedPermeability/