SNAD project, including Matvey Kornilov, associate professor at the School of Physics at the Higher School of Economics, discovered 11 anomalies, seven of which are supernova candidates. The digital images of the northern sky in 2018 were studied and searched using the nearest neighbor method based on K-dimensional trees. machine learning methods make it possible to automatically search for anomalies.
Artificial intelligence discovers new space anomaly
The research was published in the journal "New Astronomy". Most astronomical discoveries are based on observations and subsequent calculations. Back in the 20th century, the number of observations was small, but with the commissioning of wide-area astronomical surveys of the sky, the amount of data received increased many times. For example, the Zwicky Transient Facility (ZTF) is a large-scale survey of the northern sky, producing approximately 1.4 terabytes of data with nightly observations, and its catalog contains billions of objects.
Manually processing such large amounts of data is difficult and expensive, so the SNAD project team, which brought together scientists from Russia, France and the United States, tackled automating the process. To learn more about astronomical objects, scientists analyze their light curves—the dependence of an object's brightness on time.
First, a flash of light is recorded in the sky, and then how its brightness evolves: it gets brighter, weaker, or goes out completely. For the study, the scientists took the light curves of 1 million real objects from the Zwicky Transient Facility 2018 catalog and compiled seven simulated light curves of the types of objects studied. In total, around 40 properties were considered, such as object brightness and the amplitude of the periodicity.
"We characterized the properties of the simulation with a set of characteristics we would expect to see in real objects. Among a million objects we searched for super-powerful supernovae, type Ia supernovae , type II supernovae and tidal disruption events, "explains Konstantin Malanchev, one of the article's authors and a postdoc at at the University of Illinois at Urbana-Champaign. — Such objects we call exceptions. They are very rare and little research has been done on their properties, or they are interesting objects that require more detailed study. ”
Then, the light curve data of the real object is compared with the simulation using the K-tree method. A K-tree is a special geometric data structure that allows you to divide the space by using hyperplanes, planes, lines or points. Cutting into smaller parts. Splitting is used to narrow down the search in K-dimensional space. They are looking for objects with properties that are as similar as possible to those described in the seven simulations. For each of the seven simulations,
results. The 15 most similar real-life objects were found from the ZTF database. There were 105 objects in total. Their researchers manually analyzed and checked them for anomalies, and 11 anomalies were confirmed. Supernova candidates, the other four are active galactic nuclei in which tidal disruption events may have occurred
"This is a very good result," commented Maria Pruzhinskaya, one of the authors of the article and a researcher at the PK Sternberg State Institute of Astronomy. Furthermore, we not only succeeded in discovering rare objects that had already been discovered, but also discovered some new objects that had been missed by the astronomical community. This means that existing search algorithms can be debugged so that such objects are no longer missed. "
Research shows that this method does work and is very simple to implement. The proposed method of searching for specific types of celestial objects is universal and can be used to discover not only rare types of supernovae, but also other interesting astronomical objects .
"Astronomical or astrophysical phenomena not previously discovered by scientists are also anomalies," explains Matvey Kornilov, associate professor at the School of Physics at the Higher School of Economics - the observed behavior of such objects should differ from the properties of known objects in the future. It is planned to use our method to discover new object categories "
.
