Science Scribbler: Key2Cat Update from Nanoparticle Picking Workflow

Science Scribbler: Key2Cat Update from Nanoparticle Picking Workflow


This is the Science Scribbler Team with some exciting news from our latest project: Key2Cat! We have been blown away by the incredible support of this community – hundreds of you have taken part in the Key2Cat project ( and helped to pick nanoparticles in our electron microscopy images of catalyst nanoparticles. In just 1 week, over 50,000 classifications were completed on 10,000 subjects and 170,000 nanoparticles and clusters were found!

Thank you for this huge effort!

We went through the data and prepared everything for the next step: classification. Getting the central coordinates of our nanoparticles and clusters with the correct class will allow us to improve our deep learning approach. But before getting into the details of the next steps, let’s recap what has been done so far using the gold on germanium (Au/Ge) data as an example.


In the first workflow, you were asked to pick out both nanoparticles and clusters using a marking tool, which looked something like this:

As you might have realized, each of the images was only a small piece of a whole image. We tiled the images so that they wouldn’t be so overwhelming and time-consuming for an individual volunteer to work with. We also built in some overlap between the tiles so that if a nanoparticle fell on the edge in one image, it would be in the centre in another. Each tile was then shown to 5 different volunteers so that we could form a consensus on the centres of nanoparticles and clusters.


With your enormous speed, the whole Au/Ge dataset (94 full size images) was classified in just a few days! We have collected all of your marks and sorted them into their corresponding tiles. If we consider just a single tile that has been looked at by 5 volunteers, this is what the output data looks like:

With some thinking and coding we can recombine all the tiles that make up a single image, including the marks placed by all volunteers that contributed to the image:

Recontructed marked image

Wow, you all are really good at picking out the catalyst particles! Seeing how precisely all centres have been picked out in this visualisation is quite impressive. You may notice that there are more than 5 marks per nanoparticle – this is because of the overlap that we mentioned earlier. When taking the overlap into consideration, this means that each nanoparticle should be seen (at least partially!) by 20 volunteers.

The next step is to combine all of the marks to find a consensus centre point for each nanoparticle so that we have one set of coordinates to work with. There are numerous ways of doing this. One of the first that has given us good results is an unsupervised k-means algorithm [1]. This algorithm looks at all of the marks on the image and tries to find clusters of marks that are close to each other. It then joins these marks up into a single mark by finding a weighted average of their placements. You can think of it like tug-of-war where the algorithm finds the centre point because more marks are pulling it there.  

Reconstructed image with centroids of marks

As you can see, the consensus based on your marks almost perfectly points at the centres of individual nanoparticles or nanoparticle clusters. We don’t yet know from this analysis if the nanoparticle is a part of a cluster or not, and in some cases, we also get marks in areas which are not nanoparticles as shown in the orange and red boxes above. Since only small parts of the overall image were shown in the marking task, the artifact in the orange box was mistaken as a nanoparticle and in the case of the red box, there is a mark at the very edge and on a very small dot-like instance where some of you might have been suspicious about another nanoparticle. This is expected, especially since we asked volunteers to place marks if they were unsure – we wanted to capture all possible instances of nanoparticles in this first step!


This is the part where the second workflow comes into play. Using the marks from the first workflow, we createda new dataset showing just a small area around the mark to collect more information.In this workflow we ask a few questions to help identify exactly what we see at each of the marks

With this workflow, we hope to classify all the nanoparticles and clusters of both the Au/Ge and Pd/C catalyst systems, while potential false marks can be cleaned up! Once this is accomplished, we’ll have all the required inputs to improve our deep learning approach.

We’re currently collecting classifications on the Au/Ge data and will soon switch over to the Pd/C data, so if you have a few spare minutes, we would be very happy if you left some classifications in our project!

-Kevin & Michele

Got your interest? Do you have questions? Get in touch!



[1]: M. Ahmed, R. Seraj, S. M. S. Islam, Electronics (2020), 9 (8), 1295.

A Sky Full of Chocolate Sauce: Citizen Science with Aurora Zoo

by Dr. Liz MacDonald and Laura Brandt

Viewing the aurora in person is a magnificent experience, but due to location (or pesky clouds) it’s not always an option. Fortunately, citizen science projects like Aurorasaurus and Zooniverse’s Aurora Zoo make it easy to take part in aurora research from any location with an internet connection. 

The Aurorasaurus Ambassadors group was excited to celebrate Citizen Science Month by inviting Dr. Daniel Whiter of Aurora Zoo to speak at our April meeting. In this post we bring you the highlights of his presentation, which is viewable in full here

To ASK the Sky for Knowledge

Far to the north on the Norwegian island of Svalbard, three very sensitive scientific cameras gaze at a narrow patch of sky. Each camera is tuned to look for a specific wavelength of auroral light, snapping pictures at 20 or 32 frames per second. While the cameras don’t register the green or red light that aurora chasers usually photograph, the aurora dances dynamically across ASK’s images. Scientists are trying to understand more about what causes these small-scale shapes, what conditions are necessary for them to occur, and how energy is transferred from space into the Earth’s atmosphere. ASK not only sees night-time aurora, but also special “cusp aurora” that occur during the day but are only visible in extremely specific conditions (more or less from Svalbard in the winter.)

Still from Dr. Whiter’s presentation. The tiny blue square on the allsky image (a fisheye photo looking straight up) represents the field of view of the ASK cameras. The cameras point almost directly overhead. 

The setup, called Auroral Structure and Kinetics, or ASK, sometimes incorporates telescopes, similar to attaching binoculars to a camera. Project lead Dr. Daniel Whiter says, “The magnification of the telescopes is only 2x; the camera lenses themselves already provide a small field of view, equivalent to about a 280mm lens on a 35mm full frame camera. But the telescopes have a large aperture to capture lots of light, even with a small field of view.”

The challenge is that ASK has been watching the aurora for fifteen years and has amassed 180 terabytes of data. The team is too small to look through it all for the most interesting events, so they decided to ask for help from the general public. 

Visiting the Aurora Zoo

Using the Zooniverse platform, the Aurora Zoo team set up a project with which anyone can look at short clips of auroras to help highlight patterns to investigate further. The pictures are processed so that they are easier to look at. They start out black and white, but are given “false color” to help make them colorblind-friendly and easier for citizen scientists to work with. They are also sequenced into short video clips to highlight movement. To separate out pictures of clouds, the data is skimmed by the scientists each day and run through an algorithm.

Aurora Zoo participants are then asked to classify the shape, movement, and “fuzziness,” or diffuse quality, of the aurora. STEVE fans will be delighted by the humor in some of the options! For example, two of the more complex types are affectionately called “chocolate sauce” and “psychedelic kaleidoscope.” So far, Aurora Zoo citizen scientists have analyzed 7 months’ worth of data out of the approximately 80 months ASK has been actively observing aurora. Check out Dr. Whiter’s full presentation for a walkthrough on how to classify auroras, and try it out on their website!

Some of the categories into which Zooniverse volunteers classify auroral movement. Credit: Dr. Daniel Whiter.

What can be learned from Aurora Zoo is different from other citizen science projects like Aurorasaurus. For example, when several arc shapes are close to one another, they can look like a single arc to the naked eye or in a photo, but the tiny patch of sky viewed through ASK can reveal them to be separate features. These tiny details are also relevant to the study of STEVE and tiny green features in its “picket fence”.

Early (Surprising!) Results

Aurora Zoo participants blew through the most recent batch of data, and fresh data is newly available. The statistics they gathered show that different shapes and movements occur at different times of day. For example, psychedelic kaleidoscopes and chocolate sauce are more common in the evening hours. The fact that the most dynamic forms show up at night rather than in the daytime cusp aurora reveals that these forms must be connected to very active aurora on the night side of the Earth. 

Aurora Zoo participants also notice other structures. Several noted tiny structures later termed “fragmented aurora-like emissions,” or FAEs. Because of the special equipment ASK uses, the team was able to figure out that the FAEs they saw weren’t caused by usual auroral processes, but by something else. They published a paper about it, co-authored with the citizen scientists who noticed the FAEs. 

Still from Dr. Whiter’s presentation, featuring FAEs and Aurora Zoo’s first publication.

What’s next? Now that Aurora Zoo has a lot of classifications, they plan to use citizen scientists’ classifications to train a machine learning program to classify more images. They also look forward to statistical studies, and to creating new activities within Aurora Zoo like tracing certain shapes of aurora. 

STEVE fans, AuroraZoo hasn’t had a sighting yet. This makes sense, because ASK is at a higher latitude than that at which STEVE is usually seen. However, using a similar small-field technique to examine the details of STEVE has not yet been done. It might be interesting to try and could potentially yield some important insights into what causes FAEs.

Citizen Science Month, held during April of each year, encourages people to try out different projects. If you love the beautiful Northern and Southern Lights, you can help advance real aurora science by taking part in projects like Aurora Zoo and Aurorasaurus

About the authors of this blog post: Dr. Liz MacDonald and Laura Brandt lead a citizen science project called Aurorasaurus. While not a Zooniverse project, Aurorasaurus tracks auroras around the world via real-time reports by citizen scientist aurora chasers on its website and on Twitter. Aurorasaurus also conducts outreach and education across the globe, often through partnerships with local groups of enthusiasts.  Aurorasaurus is a research project that is a public-private partnership with the New Mexico Consortium supported by the National Science Foundation and NASA. Learn more about NASA citizen science here