Speaker
Description
Reservoir characterization represents a fundamental challenge in the oil and gas industry, requiring interdisciplinary integration of chemical, physical, geological, and computational analyses. Generative models have emerged as an alternative to complement real data, enabling synthetic generation of three-dimensional porous volumes with controlled properties.
This research uses an open-access microtomography dataset comprising 16 rock samples from the Brazilian pre-salt region, available in both low resolution (48 μm - 64 μm) and high resolution (6 μm - 8 μm). The dataset also includes their respective segmented images into pore and matrix.
A bottleneck in 3D image generation is the trade-off between geological representativeness and computational cost. Small subvolumes (64³ or 128³ voxels) may fail to capture key features such as long-range connectivity, fractures, or vugs, while substantially larger volumes become expensive to train and generate. Our approach employs fully convolutional architectures that keep the parameter count manageable and support variable input sizes, enabling scalable generation during inference while optimizing training efficiency.
The motivation for synthetic rock generation extends beyond data augmentation. The main goal is to enable systematic numerical experimentation under controlled conditions. By generating samples that vary in only one or two petrophysical attributes (e.g., porosity or permeability) while keeping microstructural characteristics, we can isolate individual effects and quantitatively assess their influence on transport properties like relative permeability. This capability is essential for advancing physical understanding of fluid flow in porous media.
We investigate a range of state-of-the-art generative architectures, including Generative Adversarial Networks (GANs) and diffusion models, applied to both segmented binary images and greyscale microtomography data. The framework incorporates Conditional GANs for attribute-guided generation and Wasserstein GANs for enhanced training stability. Working with segmented images enables direct modeling of pore space topology and connectivity, while greyscale generation captures the continuous attenuation characteristics and mineralogical variations inherent in raw microtomography acquisitions. Additionally, CycleGAN-based domain transfer enables cross-lithological exploration and translation between segmented and greyscale representations, facilitating investigation of how structural and textural variations influence petrophysical responses.
Ensuring physical realism requires constraints beyond simple scalar conditioning. While conditioning solely on porosity may satisfy target values numerically, it often produces topologically invalid structures with disconnected pore networks. Our approach employs Euclidean distance transform maps that efficiently encode geometric information about pore connectivity with minimal computational overhead compared to direct flow simulation.
Generated volumes are evaluated using several metrics across morphology, topology, and petrophysics. These include Minkowski functionals and spatial statistics (two-point correlation and variograms), connected component analysis and Euler characteristic, and comparisons of porosity, pore/throat size distributions, simulated permeability, and formation factor against real samples.
This integrated approach demonstrates the potential of generative models to produce geologically plausible, physically consistent synthetic rock samples that can advance reservoir characterization workflows and reduce reliance on costly experimental campaigns.
| Country | Brazil |
|---|---|
| Acceptance of the Terms & Conditions | Click here to agree |
