What is the project goal?

We need your help to find artifacts and prepare for new data from NASA's upcoming Nancy Grace Roman Space Telescope. The European Space Agency is currently flying the Euclid mission, with a camera very similar to that on Roman. Therefore, understanding artifacts that are left over from the initial data reduction from the European Space Telescope Euclid will speed up our understanding of artifacts from Roman.

Astronomers have already done a lot of data reduction work to remove bad pixels and mitigate artifacts from the images in this project -- however, it's impossible to do this perfectly! That's where you come in. We have used a machine learning model to identify artifacts in the ESA Euclid / Euclid Consortium /NASA Q1-2025 data set, but the model still makes mistakes sometimes. We can improve the model with your feedback on how well it does at characterizing artifacts.

The results from this project will be used to train further models to analyze data from NASA’s Nancy Grace Roman Space Telescope. We aim to mask out the large artifacts and be left only with data for scientific analysis to get more accurate distance measurements of distant galaxies. The final model will be publicly deployed on the Roman Research Nexus Platform for broad use by the astronomical community. Any astronomer working with Roman data will be able to access the Nexus and use our model for artifact masking.

What is an artifact?

When you take a photo with your phone, you may see false features in the image that are not really there due to dust on your camera, a lens flare, or a dead pixel. These are called artifacts. The same can happen when we take photos of the sky with telescopes! We see both the real, astronomical objects and image artifacts. However, on your phone, you can simply take another photo. With a telescope, we don’t always get that opportunity, so we have to make the best use of the image we get. Therefore, to analyze images containing artifacts, we have to understand what they look like!

Sometimes artifacts can disguise themselves as real objects. For the most accurate astronomical studies, we must be confident in which objects are artifacts and which are real objects in our universe. Each telescope has its own unique pattern of artifacts, almost like a fingerprint. When we start working with data from a new telescope, we have to figure out what that particular telescope’s image artifacts look like, so that we can obtain clean data for scientific studies. New technologies in the cameras and electronics can both lead to higher quality images and new types of image artifacts.

Galaxy Spectra from Grism Observations

The data in this project are spectroscopic observations, not images, of galaxies. Theis data can tell us about the distance to the galaxy and many of its traits. The instrument that creates this data on the telescope is called a grism.

What is a spectrum?

A spectrum is light broken up into wavelengths. You are probably familiar with this concept from a prism making a rainbow out of white light. Our eyes see different wavelengths as different colors. Knowing at which wavelengths a galaxy emits light, and how bright the light is at different wavelengths can tell us about the distance to the galaxy, the age of the stars in the galaxy, and even about its central supermassive black hole! That is because each element emits light at specific wavelengths. So we can find these chemical fingerprints in spectra, by looking for emission at those specific wavelengths.

For example, you can imagine that shining light through a prism might create the rainbow below.
(image credit: NASA)

This rainbow is currently showing light at all wavelengths, and we can’t really tell which elements might be present. However, what if we subtract the average brightness? We call that average brightness the continuum. Then we would be left with only the brightest wavelengths showing up.
(image credit: NASA)

Now, we have specific emission lines, wavelengths that are brighter than the average. This continuum subtraction step gives us a specific chemical signature to analyze and learn from.

In this workflow, you get to look at the image both before and after continuum subtraction. This can help determine if a bright spot is an artifact, an emission line, or a piece of the continuum that was imperfectly subtracted! You can imagine that each horizontal line is a rainbow, since each one is a spectrum. However, sometimes if a galaxy is faint, we only see the emission lines with no continuum, because the continuum is too dim.

The data for this project looks at the spectra of many galaxies at once, creating multiple rainbows. However, the images you see during the workflow are just one small portion of these larger images. You can imagine if the images were in color, that each spectral stripe you see would be a rainbow like in the image below.
(image credit: NASA and ESA)
The galaxy data you see will appear as streaks (continuum) or small blobs (emission lines), not as full rainbows.

Zeroth-Order Sources

One of the instruments on Roman and Euclid telescopes is called a grism. This is a combination of a grating and a prism. A grating has many tiny slits that light passes through. Like a prism creating a rainbow, the grism spreads most of the light from a distant galaxy into a spectrum, separating it by wavelength so scientists can study it. At the same time, some of the light passes straight through without being spread out.

The light that goes straight through is called a zeroth-order source. These are extra images of the galaxy created by the telescope’s optics. For the Euclid Q1 data you are seeing, these have a double blob or infinity sign shape. These double blobs are not separate galaxies, but optical artifacts created by the grism.

The light that is bent creates the spectra. These are the data that astronomers are collecting to learn about the chemical composition and distance to other galaxies.

Meta’s Segment Anything Model

The open source, Segment Anything Model (SAM) used in this project was created by Meta AI to “cut out” any object in an image. For example, see the airplane below. The SAM was trained on a variety of every day objects, so can often correctly cut out, or highlight, a new object without retraining. That’s how we can use it to identify artifacts. However, these training images were not from telescopes, so the SAM makes mistakes.
(image credits: Meta Segment Anything)
The researchers on this project will use your classifications from this workflow to improve the SAM’s training for astronomical artifact detection in grism data only. We will fine-tune our own version of the SAM's training with the results from your evaluations of the SAM's initial segmentation map predictions. We are not collaborating with Meta directly in any way. This refined, astronomical version of the SAM will be publicly released following the conclusion of the project, along with a peer reviewed journal article. We have no plans for sharing with or use by Meta, or any for-profit company. The SAM is licensed under Apache 2.0 which allows commercial use, modification, distribution, patent use and private use of the model.