Meteorite Mineral Mapping

Experiment With Technologies to Identify the Mineral Composition of Meteorites Pixel by Pixel

Hackathon Findings

• tk

Hackathon Projects

• Meteorite Mineral Mapping and its presentation
• Lapis and its presentation
• Mineral Mapping and its presentation

Background

Please note: Aspects of this challenge overlap with the Drilling Into Earth's Past challenge. Teams working on either challenge may wish to collaborate or share ideas with one another as they pursue their solutions.

Meteorites are rocks broken off from objects in the solar system that have been ejected from their parent body and landed on Earth. They can come from asteroids, comets, other planets or their moons, or any small solar system body. Meteorite strikes on the Earth are a common event, with some hundreds of them landing every year. Usually these are small, softball-sized objects. Rarely the Earth is struck by larger meteorites, such as the one that is believed to have caused an extinction event that killed the dinosaurs.

The 30,900 kg Ahnighito meteorite in the AMNH Arthur Ross Hall of Meteorites is the largest on display in any museum.

Studying the composition of meteorites tells us a great deal about the solar system and its dynamical history. The solar system began as a large cloud of gas and dust. Eventually, enough of that material coalesced to form the sun and planets, around 4.5 billion years ago. After this initial formation, there were a lot of "leftovers" that would eventually form the objects in the outer solar system, the asteroid belt, and the moons of the planets.

Understanding the composition and distribution of minerals in meteorites tells us about the composition of all solar system bodies, and allows us to predict the overall distribution of elements in our solar system. Additionally some meteorites (those that come from other planets) teach us about planetary geology too; we get the "ground truth" from meteorites as they're our primary access to actual samples from other bodies like Mars or minor planets such as 4 Vesta.

How We Study Meteorites

To analyze meteorites, we first cut a chunk the size of a shirt button from a main sample, embed it in epoxy, and polish it. The samples are then placed in an instrument called an electron microprobe (EMP). The electron microprobe laboratory at AMNH is a joint facility shared between the Museum and Columbia University's Lamont-Doherty Earth Observatory; it houses a 5-spectrometer Cameca SX-100 that was installed in July 1996. This probe can detect and analyze elements from Beryllium to Uranium at concentrations as low as 0.01 weight percent and its stage movement has a resolution of 0.5 µm (micrometers, a.k.a. microns).

The electron microprobe laboratory at AMNH with the Cameca SX-100 pictured on the far right.

The EMP channels cathode rays (a beam of electrons) from a specialized filament toward the prepared sample; when the beam hits the sample, it bounces electrons out of the elements in the sample in the form of X-rays which are emitted at a characteristic frequency for the element. These X-rays are then collected by detectors which are highly tuned to pick up on the characteristic emission frequency of a single element. The detector counts the intensity of X-rays at that frequency and converts this into an understandable percentage which represents the abundance of that element at that point in the sample. By this method we can understand the chemical make-up of a sample.

There is a limitation on this though: time. Samples are roughly 2 square centimeters in size (10,000 square micrometers) but they can contain features around tens of micrometers in size. This means that if we want to capture the full resolution of the sample, we have to have a very fine (1 micron) beam so we avoid overlapping analysis of these small features. At this resolution analysis of an entire sample becomes a task that could take weeks, thus only small areas are analyzed at a time.

Analyses can be done in two ways, each with their own advantages and drawbacks. Quantitative analyses give the exact composition of a single point in the sample but take minutes per analysis which, as discussed above, means only an extremely small portion of the sample can be analyzed. Qualitative maps can also be made of whole areas of samples, but these only return 32 bit grayscale values that represent the intensity of x-rays reaching the detector for a single element. Since these maps are obviously scaled to the highest and lowest intensity values for an element, the values could be converted to a concentration of the element, if a known intensity were also mapped. This still only gives a qualitative result though, as there are some calibrations that can not be done outside the EMP software, however because it greatly increases the number of analyses (from hundreds of single points to hundreds of thousands in a full image) it is a very valuable tool where knowing the general composition of a sample is more important than the fine details.

A sample of the 32 bit tiff files (rendered here as 8 bit) output by the probe for 10 elements in a 512 x 512 raster at 1 micron per pixel resolution. The intensity of any given pixel is related to the concentration of the given element at that pixel.

The Challenge

Making maps of objects using the EMP can tell us the relative intensity of an element at a given point in that sample, but it does not tell us what the actual mineral of that point is. Mineral identity, commonly referred to as the mineral's phase, has previously been identified in larger areas of a sample by correlating single point (quantitative) analysis, which gives the exact phase, with identifiable mineral grains in the whole map. So for example, a quantitative analysis could be done on any of the bright pixels in the image below and would give a composition of roughly 35% by weight Mg, 20% by weight Si and nothing else. In cases like this where the percentages do not add up to 100 the remainder is assumed to be O, which is often too complex to analyze individually and is therefore done by remained. This would correlate to the mineral forsterite, and we could then assume that any of the bright pixels are forsterite.

Mg map of a small area in the meteorite Semarkona. Brightest pixels can be identified as forsterite (Mg₂SiO₄) directly by quantitative analysis of a single point and assuming that all points that are similar are within error the same phase

In cases with mineral grain boundaries that are too similar to see, we can identify them chemically by combining three maps which represent the primary components of the mineral's structure (so for example, if the mineral is olivine we would use Mg, Fe, and Ca to differentiate it from something like a pyroxene or differentiate between different olivines) and combining them on the red, green, and blue (RGB) color channels as shown below.

RGB maps for Semarkona, showing how different areas are brought out based on their elemental composition. Now we are able to identify multiple different mineral phases which, if we were to conduct quantitative analysis could then be identified exactly.

The issue is that, while this approach is good for a small area or a single mineral grain, it doesn't allow for precise analysis of large areas at a time. The need to conduct time-consuming quantitative analysis on each phase is also problematic. What would be ideal is if the maps themselves could be correlated to a quantitative analysis and then compared to mineral phase to use them as a direct method of identifying minerals on a pixel-by-pixel basis, changing the number of quantitative analyses from hundreds to hundreds of thousands.

---

Solutions

As shown above we map the abundances of elements in each pixel of 2D polished slices of meteorite, producing stacks of element maps of those surfaces. Presently we use home-made software programmed in IDL that uses a decision tree to assign each pixel in the stack of maps a mineral phase based on user-defined thresholds (or combinations of thresholds) for the intensity values of elements for that phase. So for example a simple phase like pure Fe metal is easy to identify as it would have the highest gray scale intensity values of any phase in the map, and should have intensities of essentially zero for all other elements.

Where this software fails is in distinguishing phases with high complexity, or overlapping mineral compositions, as it only uses the decision tree to bin pixels, and so may misidentify a pixel if it could overlap with another (previously identified) phase. A good example of this is are the minerals olivine and pyroxene which have chemical formula (Mg, Fe)₂SiO₄ and (Mg, Fe)SiO₃ respectively. Since the amount of Mg and Fe is variable in each a low-Mg olivine could look like a high-Mg pyroxene, if the corresponding Fe percentage is not taken into account correctly. And this is only the most simplistic example.

We would like software to be able to interdependently identify mineral phase, given a set list of potential minerals with ranges of their ideal compositions. We would also like to incorporate noise (background) removal which is not currently being accounted for in our software. Ideally the end product would take a set of standards with known exact compositions, and use them to create a calibration that can convert element gray scale intensity to weight percentage of that element so that mineral phase can be identified by matching ratios, allowing the software to be independent of the data collection settings (so long as the data settings match those of the provided standards). The output of this software would be an array of pixels and their phase identity, along with a confidence interval for how well it matched with the ideal composition of that phase.

---

Resources

Data files for this challenge can be found in this challenge's challenge directory - the naming convention for the image files is as follows:

[object name]_[bit depth]_[element].tif

Here are more external links you may find useful:

• ProbeForEPMA, ProbeImage, and Cameca PeakSight software for automated data acquisition and processing. Bruker Esprit v1.8 EDS acquisition software.

• MIPAR software to try

• ENVI GIS - good lead of software to emulate (tries to figure out which pixel is a tree, which is a house - CIA uses it)

• IDL Help - check the IDL links in Online Resources And Data Sets

---

Challenge owners: Marina Gemma and Samuel Alpert