Deep Learning, Deep Images, Disk Galaxies

In the early universe, galaxies were misshapen blobs that could only dream of transforming into something as structured and stately as our present Milky Way. Unfortunately, pinning down exactly when in history some early galaxies flattened into disks is a challenging task; now, though, the formidable combination of JWST and deep learning offers steps towards an answer.

Challenging Finds

When do disky galaxies form? In theory, answering this question is straightforward: astronomers just need to look back in time, which they (somewhat remarkably) can do by looking at higher and higher redshifts, and see when disks first appear. Unfortunately but unsurprisingly, what’s simple in description is difficult in practice. High-redshift galaxies appear extremely faint in visible light and instead are brightest in the infrared, which historically was much harder to detect with high-spatial-resolution detectors. Although the Hubble Space Telescope was a transformative telescope in so many ways, it struggled to find or resolve many high-redshift galaxies.

Enter JWST. It was designed in part to find these evasive galaxies: its detectors all are sensitive to longer wavelengths of light, and especially its NIRCam instrument is expected to be a workhorse of galaxy discovery and classification. This work has already begun in earnest, and several of the observations associated with the Cosmic Evolution Early Release Science Survey (CEERS) have already been taken.

Finding Disks

Images taken by JWST of a CEERS field. Along the top and bottom are rows of inset frames which magnify specific galaxies in the image. The magnified images show generally extended disk-like shapes, but are heavily blurred/pixelated given the small angular size of these targets.

False-color JWST images and a selection of galaxies that Morpheus classified as “disky.” The photometric redshift of each galaxy is noted in the upper corner of each inset. [Robertson et al. 2023]

Though the CEERS team is still hard at work on their analysis, their data were made public immediately to allow other researchers to take a look. This allowed a collaboration led by Brant Robertson (University of California, Santa Cruz) to run their own analysis tools on a series of images taken of a region called the Extended Groth Strip.

The tool they deployed was a deep learning model named Morpheus, which back in 2020 was trained to ingest images taken by Hubble, spot high-redshift galaxies, and then classify them according to their morphology. In that earlier study, the authors compared how many galaxies were “disks” vs. “spheroids” (and other categories) at different redshifts to estimate when structure first emerged. With JWST images in hand of the exact same patch of sky, they played the same game on the same galaxies with the new data and compared the results.

A 3 row, 4 column panel figure. From left to right, the columns are labeled "Morpheus", "JWST/F150W", "Model", and "Residual". Each row corresponds to a different example galaxy, and when read across left to right demonstrates Morpheus's ability to build a model which accurately matches the data.

An example of Morpheus’s classification “under the hood.” The network assigns a probability to each pixel that it belongs to a certain galaxy classification, then assigns a classification for any object that has >50% probability of being in a certain class. [Robertson et al. 2023]

In this first-ever application of AI/ML tools to JWST data, the authors cross-matched >7,000 galaxies between the Hubble and JWST images, and interestingly, noticed that Morpheus changed its mind about how to classify some of the faintest and highest-redshift targets. Of the 160 faintest (H>24.5AB!) galaxies Morpheus claimed were disk galaxies after seeing the JWST images, it had previously only thought 5% were flat in the Hubble data. Most of the others were too faint to resolve at those shorter wavelengths, so they had previously been mislabeled as “compact,” point-like sources, when in fact there was some structure Hubble had simply missed. Excitingly, some of the newly-classified disk galaxies have photometric redshifts of 4-5, a surprisingly high value which will require further analysis to fully contextualize.

While it’s still too early for any precise estimates of the disk formation timescale, these results confirm that JWST is a transformative tool for the job. Now more than a year since launch, we can be sure to expect more galaxies, more accurate classifications, and more rewrites to our current understanding of galaxy formation in short order.

Citation

“Morpheus Reveals Distant Disk Galaxy Morphologies with JWST: The First AI/ML Analysis of JWST Images,” Brant E. Robertson et al 2023 ApJL 942 L42. doi:10.3847/2041-8213/aca086