Probabilistic Upscaling of Hydrodynamics in Geological Fractures Under Uncertainty — Data and Pretrained Model

This archive provides the large data and model assets associated with the GitHub repository https://github.com/SarPerez/Probabilistic_Upscaling_Hydrodynamics_Fractures and the article "Probabilistic Upscaling of Hydrodynamics in Geological Fractures Under Uncertainty". Fluid flow in fractured geological media is strongly controlled by aperture heterogeneity, roughness, channelisation, and uncertainty in subsurface characterisation. In natural fractures, small geometric variations can produce large changes in hydraulic behaviour, while classical deterministic aperture–permeability relationships often fail to capture the resulting variability and uncertainty. The framework associated with this archive addresses this problem through a probabilistic upscaling strategy that combines physics-based Bayesian correction, deep learning, and Darcy-scale flow upscaling. Starting from mechanical aperture fields, it predicts spatially distributed probabilistic permeability descriptors and propagates local uncertainty to effective fracture-scale hydraulic responses. The objective is to provide a scalable and physically informed alternative to deterministic permeability estimation for complex geological fractures. These large files are distributed separately from the GitHub repository because of file size constraints. The archive includes:- `dataset/Dataset_apertures_to_perms.npy`, a processed training/validation dataset used to learn the mapping from mechanical aperture patches to probabilistic permeability descriptors;- `models/residual_unet_pretrained.pth`, a pretrained Residual U-Net checkpoint used for direct inference of probabilistic permeability fields on full-fracture mechanical aperture fields. To use the archived assets after download and extraction of the GitHub repository:- place `Dataset_apertures_to_perms.npy` in the repository folder `./dataset/`;- place `residual_unet_pretrained.pth` in the repository folder `./models/`. The pretrained checkpoint is required for direct inference with `Probabilistic_Upscaling.py`. The processed dataset is required for retraining the network from scratch and for reproducing training/validation analyses. The GitHub repository contains the source code, documentation, setup instructions, and example aperture input files. This archive contains only the larger assets needed for the full reproducible workflow.