This was a detailed webinar about mineralogy via microscopy, which can be a difficult subject, in my opinion. When I was a geology undergrad, we used petrographic microscopes for optical mineralogy, analyzing shadows and grains to arrive at mineralogy. It was not my favorite class because sometimes I just couldn’t tell if the shadows were just right or not. I’m glad that the technology has improved so that machines can do most of that kind of analysis now. In my work in oil and gas, I have come across grain analysis to determine primary grains, cementation, and to attempt to unravel diagenetic alteration history. The goal here was often to understand porosity and permeability development and occlusion, to determine whether the rocks could host hydrocarbon fluids. Of course, the main use of mineralogy is probably in mineral ore analysis for mining minerals and for ore body analysis. This often involves mineral phase analysis.
This webinar was interesting
but also difficult for me since I have not used any of these newer techniques.
SEM software has been in use for 40 years, just beginning to be used when I was
an undergrad. Now. It is being replaced by new kinds of software and analysis.
Image segmentation is a new application where machine learning has excelled.
When I was an undergrad, I worked for a time in a lab with two paleobotanists.
One thing we did was to make acetate peels of Pennsylvanian fossils known as
‘coal balls’, which uniquely preserve, due to the presence of calcite, some of
the tissue structures of ancient land plants. The goal was to get a 3D analysis
of these fossils, which were inside the rock, by taking thin peels of the rock
and making thin sections that could then be stacked to get a 3D image of the
fossil. I think that would be considered a crude method of image segmentation.
Microscopy workflows are given below.
1) Multi-platform
-
2) Multi-modal
– lots of different techniques.
3) Multi-scale
microscopy
4) Multi-dimensional
1D-5D imagery. Image segmentation allows things to be measured
precisely. Electron, light, and X-ray microscopy are the means.
Deep learning neural networking is a subset of machine learning. It is based on computational algorithms. It enables better image segmentation.
Mineralogy software: SEM
software has been used for 40 years. Crushing rocks uses 5% of global
electricity. Automated mineral classification: Energy-dispersive spectroscopy
(EDS) is the main method. Automated mineralogy is an old process that is being
replaced. SEM and analysis, which is low precision (range of elemental
concentrations), is being replaced by Phase Identifier AI, which takes
segmented maps with xyz coordinates to get distributions of elements in a
sample via neural networks, from elements to minerals. It can interact with it
depending on what you are looking for. It is a data processing system. How is
the phase identifier AI unique? Image – explore – discover is the
workflow. It can be used for particle analysis, rock & core analysis,
textural analysis, concretes, factory materials, etc.
Light microscopy via a
petrographic microscope. Semantic vs. instance segmentation: Semantic is
traditional and often not very useful. Instance segmentation is far more useful
and utilizes deep learning. An example of marble sample segmentation gives a
single image vs. instance segmentation, which gives different grain sizes,
textures, and inclusions that can be analyzed. Instance segmentation can better
distinguish grains and overgrowths. It is better for shape analysis and can be
a powerful tool for geologists. Neural networks do a better job of segmentation
analysis than previous methods.
X-ray Microscopy – Zeiss
Phase identifier 3D – quantitative analysis – indirect and direct – data is
simulated in 3D. Quantitative reconstruction reduces noise. Histogram peaks are
used in reconstruction. It can yield better mineral detail. It can handle more
complex minerals. Automated mineral classification in 3D. CT scans are used in
X-ray microscopy, and one can produce moving 3D simulations.
Potential applications in
industry and academia. In industry, low concentration gold and ore body
research can benefit. It leads to better petrology.
The workflow he gives is image – explore – discover. Deep
learning happens in the explore phase.
Q&A
References? SEM info is very new, so not much.
How can the approach be validated? EDS data – can be fed in
and automatically organized. 3D and 2D comparison was done by them, not
perfect, but it validates. Dating and sequencing natural processes can be
untangled. Things like diagenesis and dissolution can be analyzed. Deep
learning allows the removal of subjective decisions. People do things somewhat
differently in their mineralogy analysis, while with deep learning, all will be
the same and more consistent.
Can it be used for clay minerals (very fine-grained)? One
can make maps finer, but it can be risky due to overlapping spots. It can be
useful for fine-grained analysis, with some limitations not really due to
software. It really depends on the resolution of the data acquired.
Sample preparation? For EDS, good data acquisition requires
good sample preparation.
No comments:
Post a Comment