Discrete Fracture Network and Pressure Transient Benchmark Dataset with Machine-Learning Workflows

This repository provides a comprehensive benchmark dataset for discrete fracture network modelling and pressure transient analysis, together with two machine-learning workflows that demonstrate how complex transient signals can be analysed and interpreted. The dataset can be used independently to develop, test, or compare new numerical, statistical, and machine learning approaches for fractured reservoirs.
The dataset was designed to represent a broad range of fractured systems. A design of experiments was used to systematically vary key fracture network properties across physically plausible, outcrop-informed ranges. Multiple stochastic realisations were generated using GeoDFN, resulting in 4,850 geologically consistent DFNs. The corresponding pressure transient dataset covers this diverse DFN ensemble under three different matrix-fracture permeability configurations (Datasets A–C), simulated using an embedded discrete fracture model (EDFM). The design of experiments, DFN generation code, and simulation scripts are all included, allowing the dataset to be regenerated or extended. Using this dataset, we developed two machine-learning workflows for characterising fractured reservoirs from pressure transient responses:Unsupervised learning of flow behaviour: The pressure transient signals (time-series data) are grouped based on similarities in their shape using Dynamic Time Warping and K-medoids clustering. The resulting clusters represent repeatable signal behaviours associated with bounded ranges of fracture network properties, while the cluster medoids provide representative responses. A random forest classifier and SHAP analysis are then used to identify which underlying physical properties most strongly control the observed signal patterns. Explainable deep learning workflow: The pressure transient signals are used directly to predict explicit discrete fracture network properties using deep learning. An attention-based interpretation step is then used to identify which parts of the signal contribute to each prediction and to assess whether the model relies on physically meaningful information. This helps address the black-box problem of deep learning by linking model predictions back to interpretable regions of the input signal. Both workflows are reusable beyond pressure transient analysis. They provide complementary approaches for time-series problems where measured signals contain information about the hidden properties or behaviour of an underlying physical system, either by identifying characteristic signal patterns without predefined labels or by directly inferring system properties while retaining physical interpretability.