Neural Networks and Space Telescope Image Analysis: Seeing the Cosmos More Clearly

NASA's first James Webb Space Telescope deep field, released July 12, 2022, covered a patch of sky the size of a grain of sand held at arm's length, and it was already crowded with galaxies. The archives behind it are far larger: Hubble has logged more than 1.7 million observations, and the Vera C. Rubin Observatory can issue up to 10 million alerts in one night. Neural networks now triage that flood. This article covers what they do, what they have found, and where astronomers still do not trust them.

Key takeaways

• Convolutional neural networks (CNNs) classify galaxy shapes, rank strong gravitational lens candidates, and separate real transient events from detector artifacts. They act as filters; people and follow-up telescopes confirm the results.
• Galaxy Zoo DECaLS paired 7.5 million volunteer classifications of 314,000 galaxies with networks that agree with confident volunteer answers about 99 percent of the time.
• In Euclid's first quick data release, a pipeline of neural networks, citizen scientists, and experts produced 497 strong-lens candidates in 63 square degrees, 250 of them top-grade.
• Image sharpening with generative networks can recover plausible detail, but it can also invent it. Test images in the landmark 2017 study were artificially degraded, not real observations.
• JWST's most distant galaxies, such as JADES-GS-z14-0, were confirmed by spectroscopy, not by machine learning. The network's job is usually to narrow the list that the spectrograph then checks.

Why telescopes produce more images than astronomers can look at

Think of an airport baggage scanner. A human reviewer can look at one bag carefully, but 100,000 bags a day need a system that flags the ones worth a second look. Telescope archives work the same way. The point is not to replace the careful look, but to decide where it should go.

NASA's "Hubble by the Numbers" page, with data as of March 2024, lists more than 1.7 million observations of about 108 million astronomical objects, an archive of about 477 terabytes, and more than 23,000 peer-reviewed papers. Hubble has been in orbit since April 24, 1990, and its 2.4 m mirror is small by today's standards. The newer facilities are far hungrier.

The Rubin Observatory's camera has about 3.2 gigapixels, built from 189 CCDs (charge-coupled devices, the light-sensing chips) of 16 megapixels each. It takes 30-second exposures, and the survey is expected to produce more than 200,000 images, or 1.28 petabytes uncompressed, per year. Full survey operations began in mid-2026, after the first images were released on June 23, 2025. The Zwicky Transient Facility (ZTF), a smaller survey at Palomar Observatory in California, covers 47 square degrees per exposure and is designed to image the entire northern sky in three nights.

Facility	Scale figure	Source
Hubble Space Telescope	More than 1.7 million observations; about 477 TB archive (March 2024)	NASA, "Hubble by the Numbers"
JWST first deep field	12.5 hours of NIRCam exposure on a patch the size of a grain of sand at arm's length	NASA, July 12, 2022 release
Zwicky Transient Facility	47 square degrees per exposure; northern sky in three nights	Wikipedia summary of ZTF design
Rubin Observatory	Up to 10 million alerts per night; alerts within 60 seconds; 1.28 PB of images per year	Wikipedia summary of Rubin design

A person could look at a few thousand of those objects in a career. The rest need software, and the software that has worked best for images is the convolutional neural network. The gap between data collected and data understood has become one of the central bottlenecks in observational astronomy, and ML is how the field is closing it.

What a neural network is doing when it looks at a galaxy

A CNN is a stack of small pattern detectors. The first layer slides tiny filters across the image, a bit like moving a stencil over a page, and each filter responds to a simple feature such as an edge or a bright spot. The next layer combines those responses into shapes like arcs and blobs, and later layers combine shapes into things a human would name, such as a bar across a galaxy's center or a ring of blue arcs around a red galaxy. The neural network architecture article explains the layers in more detail.

Most telescope applications use supervised learning. People label a set of examples, say 100,000 galaxies marked as spiral, elliptical, or merging, and the network adjusts millions of internal weights until its answers match the labels. After training, it takes a new image and outputs probabilities. A good way to picture it is a student who has gone through a huge stack of flashcards: fast and consistent on cards that look like the ones studied, and unreliable on cards from a different deck.

That last point is the source of most failures. If the training images come from one telescope and the real data from another, accuracy can drop in ways that are hard to see. Astronomers guard against this by keeping part of the data hidden during training, a rule the training and test data article explains, and by testing on objects whose answers are already known.

Sorting galaxies by shape: from Galaxy Zoo to Zoobot

Galaxy shapes carry information about history. Spirals tend to be disks with ongoing star formation, ellipticals are older and rounder, and irregular or disturbed shapes often mark a recent merger. Counting how those types change across cosmic time tests models of how galaxies grow.

Before neural networks, classifying shapes was a human job. Galaxy Zoo launched on July 11, 2007, as a volunteer project built on Sloan Digital Sky Survey images. By one account on its Wikipedia page, it collected more than 40 million classifications from more than 100,000 volunteers in about 175 days. Volunteers also spotted oddities, including Hanny's Voorwerp and the compact "Green Pea" galaxies, which closer study showed to be new kinds of astronomical object. The project grew into the Zooniverse platform, which has hosted roughly 100 citizen-science projects.

The next step was to teach machines from the volunteers' answers. In 2022, Walmsley and colleagues published the Galaxy Zoo DECaLS results in the Monthly Notices of the Royal Astronomical Society (volume 509, pages 3966 to 3988). Volunteers contributed 7.5 million classifications across 314,000 galaxies imaged by the Dark Energy Camera Legacy Survey, whose images are about 1.4 magnitudes deeper than SDSS and reveal weak bars and tidal features the older images missed. An ensemble of Bayesian CNNs was trained on those labels and, compared with confident volunteer answers, reached roughly 99 percent accuracy on every question.

Two design choices stand out. The networks output a distribution of answers, not a single label, which lets astronomers see when the network is unsure. And active learning, a method that steers volunteer effort toward the galaxies the network would learn most from, meant that about 140,000 galaxies got at least 30 volunteer classifications while the rest got roughly five. That is a sensible use of scarce human attention.

The same approach reaches JWST. In 2023, Huertas-Company and colleagues posted a morphological catalog of about 20,000 galaxies in the CEERS survey. They trained CNNs on labeled Hubble images and adapted them to JWST's NIRCam. For galaxies at redshift below 3, about 90 percent received the same early-type or late-type classification from JWST and Hubble images, and about 75 percent received the same regular-or-irregular classification. The team also reported that undisturbed disk-like systems already dominate the high-mass late-type population at redshift about 5, roughly a billion years after the Big Bang.

Project	What the network learned from	Sample	Headline result
Galaxy Zoo DECaLS (Walmsley et al., 2022)	7.5 million volunteer classifications	314,000 galaxies	About 99% agreement with confident volunteers
JWST CEERS morphology (Huertas-Company et al., 2023)	Hubble labels, adapted to JWST	About 20,000 galaxies	About 90% early/late agreement with Hubble below redshift 3
Euclid Q1 lens search (2025)	Five networks tested; the best was a fine-tuned Zoobot	About one million objects	497 lens candidates in 63 square degrees

A related project, deep learning for star-galaxy classification, covers the simpler but still important task of telling point-like stars from extended galaxies.

Sharpening images: what deconvolution and generative networks can and cannot recover

Every telescope image is a blurred version of reality. Diffraction from the finite mirror, atmospheric turbulence for ground-based instruments, and detector noise all smear the light. The blur pattern is called the point spread function, and removing it is called deconvolution. Classical deconvolution works but runs into a hard limit: where the data has no information, no algorithm can recover it.

In 2017, Schawinski and colleagues published a study in Monthly Notices of the Royal Astronomical Society: Letters asking whether a generative adversarial network (GAN) could do better. A GAN trains two networks against each other, one producing images and one trying to spot fakes, until the generator's output looks real. The GAN explainer covers the mechanics. The team trained theirs on 4,550 SDSS images of nearby galaxies, at redshifts between 0.01 and 0.02, then fed it versions degraded with worse seeing and added noise. They reported that it recovered features far better than simple deconvolution.

The result was real, and it was also narrow. The test images were degraded copies of sharp ones, so the right answer was known. The sample covered nearby galaxies from a single survey. And a generative network fills gaps with what looks plausible given its training data. It is like a restorer who has spent a career on Renaissance portraits and repairs a damaged canvas with Renaissance habits: the result looks right, and may or may not be what the painter did. For measurement science, plausible is not the same as accurate, so such sharpening is normally used to explore images, not to produce measurements.

It also helps to be precise about what happens in JWST's own pipeline. The cleanup of flicker noise (also called 1/f noise, a low-frequency pattern in detector readout) in the standard pipeline is done by conventional signal processing, using background masks, sigma clipping, and a Fourier-space or median-based noise model. It is not a neural network. Neural methods are mainly used downstream, on the finished images, by science teams.

Hunting for rare objects: gravitational lenses

A strong gravitational lens forms when a massive galaxy lies almost exactly in front of a more distant one. The foreground galaxy's gravity bends the background light into arcs or rings, magnifying it. The gravitational lensing article explains the geometry. Lenses are valuable because their bending depends on mass, including dark matter, and because the magnification lets astronomers study galaxies that would otherwise be too faint.

They are also rare, and the hunt shows both the power and the limits of neural networks. In 2017, Petrillo and colleagues applied a CNN to 255 square degrees of the Kilo-Degree Survey. Among 21,789 luminous red galaxies selected by color and brightness, three were already known lenses. The network flagged 761 candidates and recovered two of the three known lenses, and the missed one had a smaller Einstein radius than the network had been trained to see. Visual inspection by the team then cut the list to the 56 most reliable candidates. Those still needed confirmation from other telescopes.

The same pattern, scaled up, appears in Euclid's first quick data release in 2025. The discovery pipeline used neural networks for the first pass, then volunteers on the Space Warps citizen-science platform, then expert vetting and modeling. The catalog reached 497 galaxy-galaxy lenses over 63 square degrees, of which 250 were grade A, the highest confidence, and 243 of those grade A systems had not been published before. Five machine-learning models were tested, and the best performer was a fine-tuned version of Zoobot, a network pretrained on galaxy morphology. Among its top 1,000 objects ranked from about one million, 122 were grade A and 41 grade B lenses. The authors projected that the approach could yield about 7,000 high-confidence candidates by late 2026 using the survey's first full data release, and more than 100,000 across the full survey. Those are projections, not discoveries.

Search	Survey area	Network role	Outcome
Petrillo et al. (2017), KiDS	255 square degrees	CNN screen of 21,789 galaxies	761 candidates, 56 after visual inspection; 2 of 3 known lenses recovered
Euclid Q1 (2025)	63 square degrees	Five networks, best a fine-tuned Zoobot	497 lenses, 250 grade A, 243 of those new

One practical lesson is how much the volume of false positives matters. When true lenses are one in several thousand candidates, a network that is 99 percent accurate still floods the list with impostors. This is the problem the precision and recall article describes for any rare-event detector, and it is why the human vetting step stays in the pipeline.

Catching things that change: transients and real-bogus classifiers

Some of the most valuable events in astronomy last days or hours: supernova explosions, tidal disruption events where a star is torn apart by a black hole, and the afterglows of gamma-ray bursts. Finding them means comparing each night's image of the sky with a reference image and flagging what changed. Subtracting one image from another leaves thousands of candidates per night, and most are junk such as cosmic-ray hits, subtraction artifacts, and asteroids.

Sorting real from "bogus" detections is a classic job for a CNN. In 2019, Duev and colleagues described braai, a deep-learning real-bogus classifier for ZTF, in Monthly Notices of the Royal Astronomical Society (volume 489, pages 3582 to 3590). They reported low false-negative and false-positive rates, and tests on Edge TPUs, small hardware accelerators, showed comparable accuracy at lower cost. The paper also describes open-source tools for archiving ZTF alerts and for labeling training data.

Rubin raises the stakes. The observatory is designed to issue alerts within 60 seconds of an observation and up to 10 million per night. Alerts go to community "event brokers," services that add classifications and let scientists filter the stream. Brokers such as ALeRCE, Fink, and ANTARES take different approaches, and at least some use machine-learning classification scores. The point of the design is that no single team has to read the stream, so each science group can subscribe to the slice that matters to it.

A sibling application works on brightness curves rather than pictures. In a 2018 paper in The Astronomical Journal, Christopher Shallue and Andrew Vanderburg trained a CNN to tell planet transits in Kepler data from false positives. Their abstract reports that it ranks plausible planet signals above false positives 98.8 percent of the time on its test set, and it helped validate an eighth planet around Kepler-90, tying the system with our own for the largest known planetary count. The exoplanet detection article follows that story, and the LIGO machine-learning article covers a close cousin in gravitational-wave data.

JWST and the early universe: where machine learning actually sits

It is tempting to say that AI found JWST's most distant galaxies. The record is more specific. Webb's NIRCam camera finds candidates by their colors across several infrared filters, because light from the earliest galaxies is stretched, or redshifted, into the infrared. Picking promising candidates from tens of thousands of faint sources is where automated tools, including trained classifiers and photometric redshift estimators, earn their keep. Distance estimates from colors alone are rough, so the next step is spectroscopy.

The record holders were confirmed that way. JADES-GS-z14-0, found in the JADES survey, was confirmed by NIRSpec spectroscopy and reported in Nature in 2024 by Carniani and colleagues. It lies about 300 million years after the Big Bang, at a redshift of about 14.18 once later ALMA measurements refined the original NIRSpec figure of 14.32. In May 2025, JWST detected MoM-z14 at redshift 14.44, which surpassed it. Neither result depended on a neural network for its headline number. The machine learning narrowed the search, and a spectrograph made the claim.

For how Webb sees in these wavelengths, the infrared telescopes article is a useful companion, and the JWST discoveries article lists what the telescope has reported so far.

A short timeline of machine learning on telescope data

The shift from hand classification to learned classification took about a decade and happened in stages. Each row below ties to a source cited in this article.

Year	Milestone	What changed
2007	Galaxy Zoo launches (July 11)	Volunteers classify SDSS galaxies by eye; labels become training data later
2017	GAN deconvolution study; KiDS lens search	Generative networks tested on galaxy images; CNN screens 21,789 galaxies for lenses
2018	Kepler deep-learning paper	A CNN ranks planet signals above false positives 98.8% of the time on a test set
2019	braai for ZTF	A CNN separates real transients from artifacts in a live alert stream
2022	Galaxy Zoo DECaLS	Volunteer labels for 314,000 galaxies train networks with about 99% agreement
2023	JWST CEERS morphology	CNNs adapted from Hubble labels classify about 20,000 JWST galaxies
2025	Euclid Q1 lens catalog; Rubin first images	Networks, volunteers, and experts produce 497 lens candidates; Rubin's full camera releases images (June 23)

The pattern is that every row builds on labels or truth that people supplied first. The volunteers of 2007 made the networks of 2022 possible, and simulations and known lenses made the lens searches possible. Machine learning in astronomy is closer to a way of scaling up human judgment than a replacement for it.

Why a probability is not a measurement

A network's output is a number between 0 and 1 for each possible answer. It is tempting to read 0.95 for "spiral" as a 95 percent chance the galaxy is a spiral. That reading holds only if the network is calibrated, meaning that among all galaxies it scores near 0.95, about 95 percent really are spirals. Networks are often overconfident, especially on images unlike their training set.

The Galaxy Zoo DECaLS team addressed this by using an ensemble of Bayesian networks. Several networks, each slightly different, answer the same question, and the spread among them shows how uncertain the answer is. When the networks agree, the answer is more trustworthy. When they disagree, the galaxy goes into a pile for closer inspection, and that pile is often where the interesting objects are.

The same idea applies to measurements built on top of the classifications. A catalog that says "disks dominate at redshift 5" is a statistical statement about thousands of galaxies, each carrying its own uncertainty. Reporting only the best-guess label for each object throws that uncertainty away and can make a trend look cleaner than the data justify.

Where neural networks go wrong

Astronomers have developed a short list of ways a network can mislead them. Most of them have counterparts in any field that uses machine learning, and most have a standard check.

Failure	What it looks like	Common safeguard
Domain shift	A network trained on Hubble or simulated images loses accuracy on JWST or survey images	Test on a held-out real sample; compare to older results (the 90% and 75% agreement figures above)
Invented detail	A generative sharpener adds structure that looks right but was not in the data	Use for exploration, not measurement; validate on artificially degraded images with known truth
Rare-class flooding	Thousands of false lens candidates swamp a few true ones	Human vetting, spectroscopic follow-up, and reporting precision with recall
Noisy labels	Volunteers disagree, so the training answers are uncertain	Output distributions rather than single labels; weight confident answers
Selection bias	The training set misses faint or unusual objects, so the network misses them too	Check completeness with simulated objects injected into real images

Hallucination, the habit of producing fluent but wrong output, is not unique to chatbots. The AI hallucination article explains why generative models do it, and the same logic applies to image generators in the sciences. In astronomy, the cure is mostly procedural: do not use a generated pixel as a measurement, and always ask what ground truth the method was validated against.

There is also a philosophical point worth keeping in view. A classifier trained on human labels inherits human blind spots. If volunteers never labeled a type of object, the network cannot name it. Some of the most interesting finds, such as Hanny's Voorwerp, came from people noticing something that did not fit any category, and anomaly-detection methods are still catching up to that kind of surprise.

What astronomers do to check the machine

Validation in this field is layered. First comes a held-out test set with known answers, such as the three known lenses in the KiDS field, where a network that finds two of three has a measurable miss rate. Second comes comparison across instruments, as in the JWST-versus-Hubble morphology check. Third comes human review, which the Euclid pipeline builds in by routing candidates to citizen scientists and then experts.

The last and strongest layer is a different kind of measurement. A candidate lens can be confirmed by a spectrum showing two distinct redshifts along the same line of sight. A candidate high-redshift galaxy can be confirmed by identifying spectral lines. A transient can be confirmed by a follow-up telescope that sees the same event. In each case the network proposes, and an independent physical measurement disposes.

Frequently asked questions

How are neural networks used in space telescope image analysis? A: They classify galaxy shapes, rank gravitational lens candidates, separate real transient events from artifacts, and help estimate distances from colors. They act as high-speed filters over datasets too large for people to inspect, with human review and follow-up observations confirming what matters.

Can AI improve the quality of old Hubble images? A: Research methods exist that can sharpen galaxy images, such as the 2017 GAN study on SDSS data, but they are exploratory. That study tested on artificially degraded images, and generative methods can add plausible features that were not in the data. Published measurements from Hubble still rely on conventional calibration.

What is a gravitational lens, and how does AI help find them? A: A gravitational lens forms when a massive foreground galaxy bends and magnifies light from a more distant galaxy. They are rare, so networks screen large surveys and rank candidates. In Euclid's first release, a pipeline of five networks, citizen scientists, and experts produced 497 lens candidates in 63 square degrees.

Does JWST use AI in its data pipeline? A: Not for its basic detector cleanup. The standard pipeline removes flicker (1/f) noise with conventional signal-processing steps, not a neural network. Science teams use neural networks downstream, for example the CNN galaxy-morphology catalog of about 20,000 CEERS galaxies.

Will AI be used by the Vera Rubin Observatory? A: Yes, in the alert system. Rubin is designed to send alerts within 60 seconds of an observation, up to 10 million per night, to community brokers that add classifications, several using machine-learning scores. The telescope's data release and alert design are built around automated filtering.

Did AI find the most distant galaxies JWST has seen? A: No. JADES-GS-z14-0 (redshift about 14.18, roughly 300 million years after the Big Bang) and MoM-z14 (redshift 14.44) were confirmed by spectroscopy. Automated tools can help pick candidates, but the distance claim comes from spectral measurements.

The takeaway

Neural networks have changed astronomy mostly by changing where human attention goes. They do not look at the sky for us, and they do not make discoveries by themselves. They decide which of a million objects a person, a citizen-science volunteer, or a spectrograph looks at next, and the quality of that decision shapes what gets found.

The most useful way to read an AI claim in astronomy is to ask what the network replaced and what confirmed the answer. A morphology catalog replaces years of volunteer clicking and is checked against volunteers. A lens list replaces a manual search and is checked by experts. A sharpened image replaces nothing measurable and should be treated as an illustration. Where an independent measurement sits behind the claim, the result is solid. Where it does not, the claim is still a hypothesis.

Sources

• NASA, Hubble by the Numbers (data as of March 1, 2024)
• NASA, Webb's deepest infrared image of the universe yet (July 12, 2022)
• Walmsley et al., "Galaxy Zoo DECaLS," MNRAS 509, 3966 (2022), arXiv:2102.08414
• Huertas-Company et al., "Galaxy Morphology from z~6 through the eyes of JWST" (2023), arXiv:2305.02478
• Schawinski et al., "Generative Adversarial Networks recover features in astrophysical images of galaxies beyond the deconvolution limit" (2017), arXiv:1702.00403
• Petrillo et al., "Finding strong gravitational lenses in the Kilo Degree Survey with Convolutional Neural Networks" (2017), arXiv:1702.07675
• Euclid Collaboration, "Euclid Quick Data Release (Q1): The Strong Lensing Discovery Engine A," arXiv:2503.15324
• Duev et al., "Real-bogus classification for the Zwicky Transient Facility using deep learning" (2019), arXiv:1907.11259
• Shallue and Vanderburg, "Identifying Exoplanets with Deep Learning" (2018), arXiv:1712.05044
• JWST pipeline documentation, clean_flicker_noise step
• Vera C. Rubin Observatory, Zwicky Transient Facility, Galaxy Zoo, and JADES-GS-z14-0 summaries

---

Read next:

• Neural Networks and Solar Storm Prediction: AI Watches the Sun to Protect Earth
• Deep Learning and Star-Galaxy Classification: Teaching Computers to Read the Sky
• Hubble Space Telescope: The Discoveries That Changed Astronomy