Taking 3D Images of 5 nm Nanowire Bundles

An Analogy Highlighting Scale

Lets say we can think of an array of nanowires as a lawn full of green grass in suburbia. The goal is to make the perfect bushy lawn to brag to all of your suburban neighbors. So we want to make a lawn that has evenly spaced grass all of uniform height and thickness. That'll show those neighbors! Okay so lets add constraints.

Lets say we can only take pictures of our lawn from above (that is, without disturbing the grass) how do we know its local structure? What's the spacing like between blades of grass, do any of them touch? Whats the surface area of the 3D surface? Does the grass have any crap on its surface (literally)?

Okay the analogy breaks down when we realize our lawn is actually actually a little over 1 000 000 times smaller then a real lawn and its also made of delicious and nutritious Nickel.

So how do we measure these properties?

Well we start by taking the CN tower and scraping it against the lawn to cut it (the most practical way, really), we then dump all the lawn clippings into Lake Ontario. We suck up some lawn clippings (among other contaminants) from Lake Ontario using a pump, then we deposit some of these lawn clippings onto a microscope slide.

I don't mean to sound too absurdist with the analogy, but I wanted to bring to light the true absurdism in nanotechnology and nanomaterials. We are literal giants compared to what we have to handle. Bringing a scalpel blade to the surface of a nanowire coated substrate is genuinely akin to cutting your lawn with the CN tower. How do we even know if we did it right?! What if we missed and knicked a large chunk of dirt (substrate) with the grass (nanowire)? When we deposit our nanowire clippings from the scalpel into a beaker of solvent, you realize a grain of dust or sand or dirt present in the solvent, beaker surface or air is orders of magnitude larger than what we are trying to study? Its basically Lake Ontario. Then we deposit a drop of this liquid onto a transmission electron microscopy (TEM) sample holder (commonly called a TEM grid) and its ready for imaging. That is, assuming the sample is actually on it, again, impossible to determine using our eyes.

Transmission Electron Microscopy

I'm not going to get into how these beasts work, I will leave that for a later blog post.

I'm going to have to be careful since this work is unpublished. It probably will never be published considering it was just a quick characterization of one non-important sample of many samples for a collaborator, but its always wise to err on the side of caution. Its a shame academia works this way, but I respect my supervisor and our collaborators more than my storytelling.

After chopping off some of the wires above and putting them under the microscope, the first step is to determine chemical properties. Are they actually Ni wires? Are they oxidized? Another impossible thing to determine is did we oxide them by dunking them in Lake Ontario? Or when they interacted with the CN tower? Or when they were stored? So many questions...

This we will determine using Principal Components Analysis (CPA) and Blind Source Separation (a flavor of Independent Component Analysis, or ICA). A description of these methods can be found in a later blog post!

Taking Images across a Tilt Series

Okay, so our CN Tower scalpel cut a wad of grass (including some dirt) off the lawn and we managed to finagle one onto our microscope slide.

Lets add some more challenges, since its clearly not complicated enough. Imagine instead of bouncing light off our sample to image it, we can achieve a better resolution by firing bullets at it. We measure how the bullets reflect or ricochet off the sample and this is how we view our sample. You can imagine maybe the first image capture you would have an okay image even if you damage the original sample a little bit, but what if you had to take 10+ consecutive images of the same spot? It would end up substantially damaged with many bullet holes. We call this electron beam damage. There is more density of radiation at the center of an electron beam (300 keV centered on a sub-Angstrom spot) than the centre of a nuclear bomb, how crazy is that? And you expect a little wad of nanowires to handle that type of energy every time without changing? The energy is so high that it can heat the sample causing something called thermal drift, it will move away while you are imaging it causing a dragging frame effect.

I digress, lets accept that thermal drift and beam damage will happen. We then take many images at many orientations. In other words, take a picture head on, rotate the microscope slide (TEM grid) by a few degrees, then again, and again and again until a whole tilt series is captures. This is shown in the figure below.

The original image is seen on the left, an array of nanowires. Notice the areas of higher brightness and lower brightness. Since TEM is a transmission based imaging technique, thicker samples of the same element appear brighter. Bright spots mean there are overlapping wires in those locations.

Using these ~70 images we aim to reconstruct the 3D model.

The workflow is as follows

After preprocessing the image by removing spurious spikes in the image such as X-rays and cosmic rays, I also applied a band-pass filter. The reason for the band-pass filter, I will describe later.

The series of images is then cross-correlated for a rough alignment.

For a fine tuned alignment, landmarks were created on each consecutive image. If common pixels were found on multiple consecutive images (intensity based) then it was deemed a landmark.

After landmarks and cross-correlation, the image can be stitched together into a movie.

Using this aligned tilt series, we will move onto the next step of making a tomogram. A tomogram can be viewed as a projected thickness series along the z-axis. The 2D image at every step along the z-axis is a cross section. This is exactly the same of as the 2D cross sections we frequently see in CT scans or MRIs. Tomograms were created using a Simultaneous Iteration Reconstruction Technique (SIRT).

To achieve higher accuracy on the tomogram, since my original attempt was low quality, I did the alignment and preprocessing on the bandpassed image. I took the landmarks generated from the band-passed image and applied them to the original image. It was much easier to generate the landmarks from the band-passed image due to the transparancy of the image. Aligning the original image (no band-pass) using landmarks generated from itself was not accurate either.

The tomogram can be seen below.

With this aligned tomogram, it can then be converted into a segmentable model. Unfortunately due to the incredibly tough challenge of randomly oriented wires, I was not able to segment each individual wire. I simply do not have the experience in tomography for this. The mass structure can be segmented in 3D space however and interacted with. I substracted the background and did some thresholding.

The resulting 3D model is shown below. Its not perfect, and I tried many many times to improve its accuracy but sometimes you're limited by the scope of the problem. Obviously thats a cop out, but advanced reconstruction algorithms and pre-preprocessing can only go so far.