Core Image Super-Resolution for Rock Property Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for estimating pore-scale rock properties from core images are limited by the resolution of the imaging technique, leading to uncertainties in determining properties like pore-size distribution and permeability, especially in rocks with complex pore systems, as high-resolution imaging is time-consuming and typically done on small samples that may not represent formation heterogeneity.
Innovation Solution
A method is developed to refine estimated parameter values within core images using a trained model, which generates a conditioned core model by aligning and enhancing low-resolution images to match high-resolution images, allowing for accurate estimation of rock properties from larger rock volumes using multi-layer convolutional models, generative adversarial networks, or U-Net variants.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If high-resolution imaging is used to capture pore-scale details, then measurement precision of pore properties is improved, but productivity decreases due to time-consuming imaging process
Solution Approach 1:
The patent applies preliminary action by pre-training a deep learning model using high-resolution images before actual analysis. The model learns to map low-resolution features to high-resolution pore characteristics during the training phase, enabling rapid inference on new samples without requiring time-consuming high-resolution imaging of each sample.
Solution Approach 2:
The patent uses copying by generating synthetic high-resolution images from low-resolution inputs through the trained neural network model. The model creates virtual high-resolution representations that replicate the pore-scale details normally obtained through expensive and time-consuming physical high-resolution imaging, allowing rapid analysis without actual high-resolution scanning.
2Measurement precision
If high-resolution imaging is performed on small samples to achieve detailed pore characterization, then measurement precision is improved, but reliability decreases due to insufficient representation of formation heterogeneity
Solution Approach 1:
The patent applies dimensionality change by transitioning from 2D image resolution enhancement to 3D volume reconstruction. The model processes 2D low-resolution images and generates corresponding 3D pore-scale models, enabling volumetric analysis that better represents formation heterogeneity while maintaining pore-scale accuracy through the learned mappings.
Solution Approach 2:
The patent uses parameter changes by transforming the resolution parameter from low to high through the neural network transformation. The model learns to predict high-resolution pore structure parameters (pore size, shape, connectivity) from low-resolution input parameters, effectively changing the resolution parameter without requiring physical re-imaging at higher resolution.
3Productivity
If low-resolution imaging is used to image large rock volumes, then productivity is improved by covering larger areas, but measurement precision of pore properties deteriorates
Solution Approach 1:
The patent introduces an intermediary - a deep learning-based super-resolution model - that bridges the gap between low-resolution input images and high-resolution pore property estimates. The model acts as a mediator that transforms low-resolution image data into high-fidelity pore-scale predictions, enabling accurate property estimation from low-resolution large-volume images without requiring actual high-resolution scanning.
Data Source
AI summary
A method is described for training a model that refines estimated parameter values within core images is disclosed. The method includes receiving multiple training image pairs wherein each training image pair includes: (i) an unrefined core image of a rock sample to be used for estimating rock properties, and (ii) a refined core image of the same rock sample; generating a training dataset from the multiple training image pairs; receiving an initial core model; generating a conditioned core model by training, using the multiple training image pairs, the initial core model; and storing the conditioned core model in electronic storage. The conditioned core model may be applied to an initial target core image data set to generate a refined target sore image dataset. The method may be executed by a computer system.