GAN River-I: A process-based low NTG meandering reservoir model dataset for machine learning studies

Area of Science:

• Geosciences
• Machine Learning
• Reservoir Engineering

Background:

• Realistic facies distributions are crucial for accurate subsurface reservoir modeling.
• Existing datasets often lack the complexity and non-stationarity needed to test advanced geostatistical and machine learning tools.
• Meandering fluvial systems present complex geological structures that are challenging to replicate.

Purpose of the Study:

• To introduce the GAN River-I dataset, a novel resource for evaluating machine learning and geostatistical models.
• To provide a benchmark for generative models aiming to recreate complex subsurface reservoir facies distributions.
• To facilitate the development and comparison of tools for reservoir characterization.

Main Methods:

• Generation of 25 3D facies models using a process-based simulator of a meandering fluvial system.
• Conversion of 3D models into three distinct datasets comprising 16,000 2D models/images, each with decreasing facies complexity (9, 7, and 3 facies).
• Amalgamation of similar facies based on permeability and sedimentary relationships to represent reservoir flow units.

Main Results:

• The GAN River-I dataset features non-stationary facies distributions with complex geometries, exceeding the complexity of previous open datasets.
• The dataset includes varying numbers of facies (9, 7, 3) across three distinct sets, allowing for controlled complexity analysis.
• Ensembles of meandering models with varied avulsion rates are provided, suitable for benchmarking generative models.

Conclusions:

• GAN River-I serves as a valuable and geologically plausible benchmark for testing and advancing machine learning and geostatistical methodologies in reservoir modeling.
• The dataset's structured complexity enables researchers to systematically assess and improve their models' ability to replicate realistic subsurface heterogeneity.
• Availability in multiple formats (image, Ndarray, GSLIB) enhances accessibility for diverse research needs in subsurface characterization.