Grain-Based Mineralogy Segmentation With Machine-Learned EDS Enhancement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Automated mineralogical analysis of complex geological samples is challenging due to the difficulty in accurately segmenting and classifying touching particles, which is exacerbated by the limitations of histogram-based thresholding methods and the lower resolution of Energy Dispersive X-ray Spectroscopy (EDS) images, leading to unreliable results and increased costs.
Innovation Solution
Employing machine learning techniques for image segmentation and classification, including unsupervised clustering algorithms, instance segmentation, and convolutional neural networks to enhance EDS image resolution using data from higher resolution modalities like BSE, thereby enabling accurate separation and classification of grains in mineralogical samples.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If histogram-based thresholding methods are used for particle segmentation, then the process is simple and fast, but the segmentation accuracy deteriorates for complex samples with touching or clumped particles
Solution Approach 1:
The patent replaces traditional histogram-based thresholding methods (mechanical/image processing approach) with machine learning-based segmentation algorithms. The ML model learns complex patterns and boundaries from training data, enabling accurate segmentation of touching and clumped particles that traditional methods cannot separate, thus improving measurement precision while maintaining automated processing efficiency.
2Loss of information
If EDS imaging is used for mineralogical analysis, then compositional information is obtained, but the image resolution is lower compared to other modalities
Solution Approach 1:
The patent merges EDS imaging data with higher-resolution imaging modalities (such as BSE or optical microscopy) to create a composite analysis system. The high-resolution modality provides detailed morphological information while EDS provides compositional data, and the machine learning model integrates both data sources to achieve both high resolution and compositional accuracy simultaneously.
3Measurement precision
If manual point counting is performed by skilled petrographers, then accurate mineralogical analysis is achieved, but the process is slow, laborious, and expensive
Solution Approach 1:
The patent implements an automated machine learning system that performs mineralogical analysis independently without requiring skilled petrographers to manually count points. The system trains on labeled data and then autonomously segments and classifies particles in new samples, achieving both high accuracy and rapid processing of entire rock core lengths, thereby eliminating the labor-intensive manual process.
4Ease of manufacture
If traditional segmentation methods are used for simple samples, then the analysis is straightforward, but they fail to provide reliable results for complex, convoluted, and diverse samples
Solution Approach 1:
The patent employs dynamic machine learning models that can adapt their complexity and parameters based on the characteristics of the input sample. For simple samples, the system can use faster, simpler algorithms, while for complex, convoluted, and diverse samples, it automatically switches to more sophisticated segmentation strategies with higher computational power, ensuring reliable results across all sample types.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A method of enhancing a resolution of an EDS image of a sample includes generating an EDS image of the sample, generating a non-EDS image of the sample generating, using a machine learning algorithm, an enhanced resolution EDS image of the sample based on the generated feature map and based on the first EDS image, where a resolution of the enhanced resolution EDS image is higher than a resolution of the first EDS image.