Art Courtesy of Madeleine Popofsky
Machine learning (ML) algorithms are addressing one of the biggest challenges astronomers currently face: digging through the vast mountains of data they have collected. “We are now in the epoch of big data in astronomy, where there is more data than there are astronomers to process it,” said Samuel Lai, an astronomy PhD student at Australian National University. Such a data surplus makes ML algorithms that automatically process data especially valuable in astronomical research. These algorithms can comb through data collected from telescopes to finish tasks like classifying astronomical objects that would take astronomers much longer to complete by hand. “Although machine learning has been used for decades in astronomy, in the grand scheme of things, we’re only just starting to explore its applications,” Lai said.
Though promising, ML algorithms have a critical flaw: they are dependent on their training data. ML algorithms “learn” by processing training data, which humans first categorize to develop the algorithm’s ability to recognize patterns and extrapolate new data. ML algorithms often fail when they encounter extreme cases in datasets, since exceptional cases are, by their very nature, outside of the patterns that the algorithms use to make sense of large pools of data. A striking example of this failure has emerged in the case of a record-breaking quasar: the newly identified J0529-4351. This astronomical object, first spotted in the sky by astronomers in 1980, has only recently been recognized as the most luminous object in the known universe, according to a paper published in Nature Astronomy.
Quasars, short for quasi-stellar objects, are often mistaken for stars near Earth because of their remarkable brightness. However, quasars are not stars; they are actually black holes in distant galaxies that heat their surroundings so much that everything around them starts to glow. As gravity sucks the matter around the black hole inward and energy is transformed into heat, the matter immediately surrounding the black hole emits high-intensity light. “The light from this glow escapes and travels through incredible distances across the Universe to reach us and our telescopes,” Lai said. A quasar in another galaxy might rival the brightness of a nearby Milky Way star when viewed from Earth. While the relative brightnesses are similar, the objective measure of light emission known as luminance shows dramatic differences.
The particular quasar known as J0529-4351 is one of many discovered by Lai and other researchers at the Research School of Astronomy and Astrophysics at the Australian National University based on recent computational work. The team’s initial goal was to identify all the quasars in the sky visible to the space telescope Gaia by combining observational data from Gaia and another telescope’s sky survey. First, they filtered out objects that didn’t exhibit quasars’ characteristic lack of movement. Astronomers cannot perceive the movement of quasars due to their great distance from Earth, but they can for closer objects such as Milky Way stars. The researchers also filtered out objects that lacked a particular effect of viewing known as parallax, in which the position or direction of the object appears different when viewed from different points in Earth’s orbit. Second, they used mid-infrared photometry, a technique of analyzing light, to discriminate between the spectrum of colors emitted by quasars and other objects. Among the objects that made it through the screening was J0529-4351.
The data from Gaia were originally published in 2022 and included a color spectrum of J0529-4351 that a human astronomer would easily detect as derived from a quasar. However, when these data were initially analyzed by a machine-learning algorithm known as the Gaia Discrete Source Classifier, the quasar was given a 99.98 percent probability of being another star in the Milky Way.
“For quasars outside the training sample, especially record-breaking quasars like J0529-4351, the models may be led to believe that the target has attributes of both stars and quasars,” Lai said. “Depending on how those attributes are weighted, the machine learning model can fail to characterize the object accurately, and we have a concrete example for J0529-4351.”
Based on the manual screening process, the researchers produced “The All-sky BRIght, Complete Quasar Survey,” a comprehensive index of quasars visible to Gaia. In addition to having near-perfect identification of quasars already documented, it was also highly accurate in identifying more than one hundred new quasars, as measured by comparison with other recent astronomical research. Using the survey, the researchers took note of J0529-4351 and then began learning more about its unique luminosity.
Typically, more luminous quasars indicate larger black holes. By measuring the speed at which matter orbited J0529-4351, Lai and his colleagues were able to calculate the mass of the black hole to be seventeen billion times the mass of the Sun. Beyond just that, they found that it is the fastest-growing black hole discovered in the Universe, as shown by data from the Very Large Telescope at the European Southern Observatory. “To shine with the measured luminosity, the black hole is possibly feeding material at a rate of one Sun mass every day, or equivalently, about four times the mass of the Earth every single second,” Lai said.
Ultramassive black holes such as this one present a scientific mystery. Because they are so far away, the light from these quasars is from quite early in the Universe; the light from J0529-4351 took twelve billion years to reach Earth. “At such an early time in the Universe, we [thought] that there hadn’t been enough time for the black holes to grow to those sizes under the normal processes that we know of,” Lai said. Discovering such a luminous quasar provides invaluable data that can be used to answer further questions about how black holes grow as astronomy advances. In another intriguing future project, these quasars might be used to directly measure the expansion of the Universe through continual observation over many years.
Despite the failure of machine learning in the case of J0529-4351, Lai was optimistic about its future use to astronomers. “By combining known physics with machine learning, it will help the statistical models develop meaningful and physically motivated correlations, which should help achieve more accurate and reliable predictions,” Lai said. To discover quasars, scientists have historically changed methodologies many times. In the 1960s, research was done using radio detection to separate quasars from stars, which have very different appearances to radio telescopes. Later surveys used methods such as color spectrum analyses and distance measurements. In the new age of machine learning, astronomers have yet another chance to refine their techniques.
The researchers contend that it is unlikely that a quasar brighter than J0529-4351 will be found. However, there is a possibility that such quasars could exist if they are located in what is known as the galactic plane, an area with a high density of matter from the Milky Way within which identification of quasars is uniquely difficult because of interference. Only with further searching—and a human touch to catch the most surprising cases—can the study of quasars advance.