Chinese AI Classifies More Than 27 Million Stars, Galaxies and Quasars

Two Types of Data Help Sort Millions of Cosmic Objects

Scientists at China's Yunnan Observatories have developed an artificial intelligence model that has classified more than 27 million celestial objects across about 1,350 square degrees of sky. By examining both how an object looks and how its light is distributed across different wavelengths, the neural network separates stars, galaxies and quasars on a scale that traditional methods struggle to match.

The team trained the model using sources whose identities had been confirmed through spectroscopy in the Sloan Digital Sky Survey's 17th data release. It then applied the system to images from the Kilo-Degree Survey, known as KiDS, focusing on objects brighter than r = 23 magnitude. That threshold sets the brightness limit for the objects included in the classification.

Checks against other astronomical surveys produced high agreement rates. Among 3.4 million sources from the Gaia mission, primarily stars with well known motions, the model identified 99.7% as stars. Validation using the Galaxy And Mass Assembly survey, or GAMA, also achieved a 99.7% success rate for galaxies and quasars.

The work addresses a practical problem for astronomy: telescopes can gather images much faster than scientists can examine and label everything in them. A useful catalog must do more than record bright dots. It must tell researchers which dots belong to different kinds of objects so they can investigate the universe's structure and history.

As of July 28, 2025, the Yunnan project sits within a wider shift toward combining artificial intelligence, professional astronomy and public participation. Its particular contribution is to bring two kinds of evidence into the same classification process, helping the model distinguish objects that can look very similar in survey images.

Why Telescope Images Are Difficult to Classify

The Sloan Digital Sky Survey, KiDS and the European Space Agency's Euclid mission are collecting images and measurements of billions of objects. Each new survey adds terabytes of data, recording galaxies billions of light-years away, supernovae and the large scale cosmic web that connects structures across the universe.

Astronomers have traditionally relied on spectroscopy to determine an object's composition, distance and type. Spectroscopy examines light in detail, giving scientists information that an ordinary image cannot provide. Its precision makes it valuable for confirming classifications, but obtaining and analysing spectra takes time and resources.

Photometric surveys offer a faster way to collect information from large areas of sky. They measure objects in different colours, or wavelengths, rather than gathering the full detail of a spectrum. This makes them useful for building large catalogs, although their measurements can leave an object's identity uncertain.

Distant galaxies, stars and quasars may all appear as small, bright points. Shape alone can therefore be a poor guide to what an image contains. Colour provides another clue, but the Yunnan approach is built around the idea that appearance and light measurements are more useful when considered together.

Human inspection and methods that require substantial manual involvement cannot keep pace with the growing volume of observations. The challenge is also scientific: faint objects, distant objects and objects seen through cosmic dust can be harder to classify, even when a computer is doing the initial sorting.

How the Yunnan Neural Network Works

The Yunnan model processes morphology and spectral energy distribution information at the same time. Morphology means an object's visible form or appearance. Its spectral energy distribution, often shortened to SED, describes how its light is spread across wavelengths. Together, these give the system both visual and colour information.

Earlier approaches have often relied on one of these types of evidence. A model looking mainly at appearance may struggle when several kinds of objects resemble points of light. A model using light measurements alone leaves out potentially useful information about shape. The new design allows the neural network to use subtle differences in both.

The Sloan Digital Sky Survey training sources gave the team a reference set with classifications confirmed by spectroscopy. Training against confirmed examples lets the model learn patterns associated with stars, galaxies and quasars before applying those patterns to the much larger KiDS sample.

The resulting classification covered more than 27 million objects within the stated brightness limit and sky area. The value of that scale is practical: researchers can work with a broad, organised catalog rather than face millions of images whose contents still need an initial label.

What the Checks Against Gaia and GAMA Showed

Gaia supplied a large check on the model's identification of stars. The validation involved 3.4 million sources, primarily stars with well known motions, and the model classified 99.7% as stars. This gave the team a way to compare its output with information from another major mission.

GAMA provided another test, covering the classification of galaxies and quasars. The model achieved a similar success rate of 99.7%. Testing against more than one survey matters because the system is intended to be useful beyond the particular examples on which it learned.

The neural network also corrected some earlier catalog mistakes, reclassifying objects that had previously been assigned the wrong label. Catalog maintenance is a substantial part of the task: researchers need reliable collections of objects as well as discoveries of new or unusual ones.

The two validation exercises give astronomers concrete measures of performance for those checks. Reliability across a wider range of observations also depends on the conditions in which the model is used, including the brightness of objects, the instruments collecting the data and the kinds of examples represented during training.

Euclid and Other Surveys Increase the Workload

Euclid illustrates how quickly the amount of astronomical data is growing. The European Space Agency mission has already mapped 26 million galaxies in only a fraction of its planned survey area. Over the next six years, it is expected to image more than 1.5 billion galaxies and send back around 100 gigabytes of data each day.

The upcoming Vera Rubin Observatory is expected to detect millions of new supernovae and other transient events. A transient is an event that appears or changes over time, so finding it and determining what it is adds another kind of workload to the task of organising survey data.

Artificial intelligence is already being used to automate parts of that process. The Bright Transient Survey Bot, known as BTSbot, supports automated discovery and classification of supernovae in the Zwicky Transient Facility survey. It is an example of AI being applied to changing events as well as to large collections of stars and galaxies.

The Yunnan model and these other tools address related needs, but their jobs vary. Sorting stars, galaxies and quasars builds a catalog of object types. Identifying supernovae helps researchers recognise events. Examining galaxy shapes provides information about structure. Modern surveys create demand for all of these forms of analysis.

Volunteers Help AI Handle Complex Galaxy Images

Public participation remains part of astronomical classification. Projects such as Galaxy Zoo invite volunteers around the world to examine galaxy shapes. Galaxy Zoo: Euclid extends that approach to images from the Euclid mission.

These projects pair human judgement with artificial intelligence. ZooBot handles simpler cases and passes more complex images to people. Volunteers can concentrate on galaxies that need closer attention, while the computer processes the easier examples at a much greater scale.

Human classifications also help train and improve the AI. The relationship works in both directions: automated sorting reduces the amount of routine inspection, and the decisions made by volunteers provide information that can improve later classifications.

NASA's James Webb Space Telescope and the Euclid mission both rely on this combination of volunteers and advanced AI in work to build some of the largest and most detailed catalogs of galaxy morphology. Thousands of participants contribute to the effort, making the organisation of astronomical images a task shared beyond professional research teams.

An internet connection gives members of the public a way to take part in mapping the cosmos. Their role is especially useful when an image is ambiguous or contains a complex structure that needs human attention.

From Better Catalogs to Rare Ring Galaxies

The GALAXY CRUISE project offers a concrete example of what AI and citizen science can achieve together. Their combined work identified 30,000 rare ring galaxies, creating a sample large enough for meaningful scientific analysis of these unusual objects.

Ring galaxies often form through collisions involving spiral galaxies. Finding a large collection lets researchers investigate how those structures develop, moving beyond the identification of individual striking images toward the study of a broader population.

Fast classification also helps researchers search for distant quasars and other rare or unusual objects that might otherwise be overlooked in enormous catalogs. The same tools can improve catalog completeness, fill gaps and correct earlier mistakes.

Galaxy classifications support studies of structure and evolution across cosmic time. Researchers can use organised collections to examine how galaxies have changed over billions of years and how their distribution fits into the cosmic web.

Dark matter and dark energy research also benefits from measurements of galaxy shapes and positions. Gravitational lensing, the distortion of galaxy images associated with gravity, provides information that astronomers can analyse when mapping the universe. Classification is part of the wider process of turning telescope observations into material suitable for those investigations.

New Models Combine Images and Measurements

Other advances in astronomical AI follow a similar principle to the Yunnan system: combine different kinds of information rather than depend on a single view of an object. Multi-modal deep learning models bring image data together with tabular measurements such as brightness, colour and redshift.

Studies of these approaches have found better precision and recall than models using only one type of data, with fewer training epochs. Precision concerns how reliably a model assigns a label; recall concerns how successfully it finds the objects belonging to that category. Training epochs are repeated passes through the training data.

Morpheus, an AI tool developed at the University of California, Santa Cruz, takes another approach by analysing astronomical images pixel by pixel. This allows detailed morphological classification and helps identify complex structures within galaxies, rather than assigning only a single broad description to an entire image.

Euclid's planned work requires attention to particularly subtle features. The mission aims to map the three dimensional distribution of dark matter by measuring small distortions in galaxy shapes caused by gravitational lensing. Processing those measurements across vast numbers of galaxies requires advanced AI and machine learning techniques.

These developments show why astronomical AI includes several kinds of tools. Some sort whole objects into categories, some combine images with measurements, and some examine the structure within an image. Each addresses a different part of the work created by increasingly ambitious surveys.

Dust, Faint Objects and Training Data Remain Challenges

Research from the University of La Serena in Chile highlights the difficulty of classifying objects in the zone of avoidance, a region where dust blocks much of the visible light. Machine learning models can produce high error rates there unless their training includes diverse data gathered across multiple wavelengths.

Models can also struggle with faint or distant objects and with sources that have unusual properties. In difficult environments, they may fail to detect or correctly identify a substantial fraction of objects. The quality and variety of training examples therefore have a direct effect on the usefulness of the resulting catalog.

Decision trees and random forests, both established machine learning methods, have been used for astronomical classification alongside more recent deep learning neural networks. These approaches have achieved strong accuracy, yet the difficult cases remain a concern across automated methods.

A model trained on one survey may perform less well with another instrument or in another region of sky. Differences in observations and in the objects being examined can change how well the patterns learned during training apply to new data.

Training data can also pass biases into an AI system, causing it to overlook rare or unexpected phenomena. Professional astronomers and citizen scientists remain important for checking results and recognising objects that require more than a routine label.

The Next Work Is to Reach Fainter Objects

The Yunnan team and other researchers are working to improve classification of even fainter objects and to adapt models to surveys with different characteristics. Their approaches include using more diverse training datasets and bringing together measurements from multiple wavelengths.

Researchers are also developing algorithms that can flag uncertain or unusual cases for human review. This makes the allocation of attention part of the classification process: routine cases can be processed automatically, while the objects that are hardest to interpret receive closer examination.

Automating the repetitive work gives astronomers more time for interpretation, theory and the search for unfamiliar phenomena. The contribution is larger than speed alone. A more accurate and complete catalog changes which objects researchers can select for study and how confidently they can compare populations.

Euclid, James Webb and the upcoming Vera Rubin Observatory are set to deliver enormous volumes of observations. The Yunnan result provides a concrete example of how AI can organise that information, while work by volunteers and professional astronomers supplies the judgement needed for complex cases and new discoveries.

Key Points

• Yunnan Observatories developed a neural network that classified more than 27 million stars, galaxies and quasars across about 1,350 square degrees of sky.
• The model combines appearance and spectral energy distribution information, using training sources from the Sloan Digital Sky Survey's 17th data release.
• The KiDS classification focused on objects brighter than r = 23 magnitude.
• Validation identified 99.7% of 3.4 million Gaia sources as stars and achieved a 99.7% success rate for galaxies and quasars in GAMA.
• Euclid has mapped 26 million galaxies and is expected to image more than 1.5 billion over the next six years.
• Galaxy Zoo, ZooBot and other projects combine automated processing with human inspection; GALAXY CRUISE identified 30,000 rare ring galaxies.
• Dust, faint objects, unusual sources and differences between surveys remain challenges, with further work focused on broader training data and human review of uncertain cases.