Speaker
Description
The digital characterization of porous media is undergoing a profound transformation driven by Artificial Intelligence (AI). However, the adoption of deep learning in Digital Rock Physics (DRP) is often hindered by the fragmentation of scientific workflows requiring separate, disconnected tools for image visualization, data annotation, and model training. We present GeoSlicer, an open-source, multi-platform software based on the robust 3D Slicer architecture, designed to unify these critical tasks into a single, cohesive environment. GeoSlicer democratizes access to advanced AI by bundling industry-standard deep learning frameworks, including TensorFlow and PyTorch, directly within its Python environment. This integration eliminates the complex dependency management that typically challenges geoscientists, enabling the seamless deployment of neural networks for reservoir characterization.
GeoSlicer excels as a comprehensive workbench for machine learning data preparation, addressing the "ground truth" bottleneck that limits supervised learning. It offers a suite of advanced annotation tools, allowing users to rapidly generate high-quality semantic labels for 3D micro-CT and thin-section imagery. Features such as semi-automated segmentation (e.g., fast marching, region growing), logical masking, and interactive thresholding streamline the creation of training datasets. Once annotated, data can be efficiently processed using internal pipelines that leverage HDF5 and out-of-core handling of massive volumes (e.g., $3000^{3}$ voxels), ensuring that multiscale data,from microCT, coreCT, well logs and thin sections,can be analyzed on standard workstations. The platform further supports real-time training monitoring via integrated TensorBoard visualization, closing the loop between geological interpretation and model performance.
In the context of AI for generating multiscale images, integrating microCT and coreCT data, for example, we modified the SinGAN (Single Image Generative Adversarial Network) model by integrating 3D convolutional layers, enabling it to process volumetric data. To address the memory constraints inherent in the original architecture, we developed Early Cropping and Patched Inference techniques, enabling generating images of $10^{10}$ voxels. We have named this 3D rock generation model as RockSinGAN, which was integrated into the GeoSlicer ecosystem, marking a significant leap in digital rock generation. Unlike traditional deep learning models that require thousands of training examples, RockSinGAN allows for the training of a generative model using a single representative 3D reference image. This capability enables the synthetic generation of large, statistically equivalent 3D rock volumes from limited input data. The model has a pyramidal resolution architecture which allows the integration of rock images in different scales as conditioning data. By generating stochastic realizations of the pore structure, RockSinGAN facilitates rigorous multiscale analysis and uncertainty quantification, providing researchers with a new tool to assess the impact of heterogeneity on rock properties essential in reservoir models.
| Country | Brazil |
|---|---|
| Acceptance of the Terms & Conditions | Click here to agree |
