Collection of Porous Media Datasets for Permeability Prediction (with Deep Learning)
Description
This dataset contains three collections of porous media geometries and corresponding flow simulation data designed for training and evaluating Fourier Neural Operators (FNOs) for permeability prediction. The data supports the research presented in "Estimating the Permeability of Porous Media with Fourier Neural Operators" and enables physics-informed machine learning approaches to fluid flow prediction in complex geometries. The corresponding code is availibe on GitHub (https://github.com/NuxNux7/Permeability-from-FNO) Dataset Collections 1. Shifted Spheres Dataset (3D) - 280 samples of structured sphere geometries in 256×128×128 voxel domains - Spheres with diameters ranging from 15-30 lattice units - Random positional shifts applied to increase geometric complexity - Achieves Reynolds numbers around 0.05 for creeping flow conditions - Serves as a controlled benchmark for model validation 2. Digital Rocks Portal Dataset (3D) - 119 successfully simulated complex geometries from front and back of diverse rock types - Original 256³ geometries scaled to 128³ with 32-cell buffer zones - Sourced from the Digital Rocks Portal under ODC Attribution License (E. Santos, J., Chang, B., Kang, Q., Viswanathan, H., Lubbers, N., Gigliotti, A., Prodanovic, M., 2021. 3d dataset of simulations. https://www.doi.org/10.17612/93pd-y471) - Represents realistic porous media with varying connectivity and tortuosity - Includes challenging cases with narrow connecting channels 3. Sandstone Dataset (2D) - ~1000 smooth sandstone images with curved Voronoi polygon structures - ~500 rough sandstone samples with increased surface complexity - ~200 rock images for additional geometric diversity - 512² resolution with 64-cell buffer zones for boundary conditions - Sourced from paper: Geng, S., Zhai, S., Li, C., 2024. Swin transformer based transfer learning model for predicting porous media permeability from 2d images. Computers and Geotechnics 168, 106–177. URL: http://dx.doi.org/10.1016/j.compgeo.2024.106177, doi:10.1016/j.compgeo.2024.106177) Data Format and Structure All datasets are provided in HDF5 format for efficient storage and access. Each file contains: - Scaled Geometry data: Smoothed porous media structures (input) - Scaled Pressure fields: Complete 2D/3D pressure distributions from lattice Boltzmann simulations (target) - Boundaries: Used for scaling [offset, scale, pow, min_pow, max_pow] - Names Spliting in train and validation - Due to a special sorting avoiding rotated geometries and the front to back inverted simulations in both sets, only the sorted datasets are present - Feel free to contact if there is intrest in the unaltered simulation data.
Files
Steps to reproduce
Flow simulations were conducted using the lattice Boltzmann method (LBM) with: - Two-relaxation-time collision operator optimized for porous media - Velocity boundary conditions (inlet) and pressure boundary conditions (outlet) - Steady-state convergence criteria based on L2 norm of pressure fields - Reynolds numbers maintained below 0.1 for laminar flow assumptions 1st dataset was simulated in waLBerla for a performance benchmark from Samuel Kemmler et al. 2st and 3rd dataset were simulated using lbmpy. The code is availible at GitHub (https://github.com/NuxNux7/Permeability-from-FNO/tree/main/lbmpy) Preprocessing and Normalization The data has been pre-scaled and normalized for immediate use in machine learning workflows: - Pressure fields: Normalized to range with moderate clipping for outliers - Geometry data: Anti-aliased and smoothed binary fields converted to range - Exponential transformation: Applied to datasets 2 and 3 (power of 0.5) to address skewed pressure distributions - Augmentation: 90° rotations applied to 3D geometries (4× data multiplication) and front to back flow in dataset 2 to increase training data