Enabling New Discoveries with Machine Learning

The universe is becoming too loud for human ears to listen to alone. In the coming years, massive new telescopes will begin scanning the sky, capturing images and signals at a pace that dwarfs anything astronomers have ever faced. One project, the Vera C. Rubin Observatory, will eventually catalog roughly 20 billion optical galaxies and will spot something changing in the sky more than 10 million times every single night. Another, the Square Kilometre Array, will discover over a billion radio galaxies. When the data arrives in such overwhelming quantities, the old method of a scientist sitting down to look at every single image or signal one by one becomes impossible. The challenge is no longer just gathering the data, but finding the needle in the haystack when the haystack is the size of a planet.

To solve this, researchers are turning to machine learning, a field where computers learn to recognize patterns without being explicitly programmed for every rule. However, a simple computer program that flags anything unusual is not enough. In a dataset filled with billions of objects, a computer might flag a glitch in the camera or a smudge on a lens as "strange," even though it holds no scientific value. The real goal is to find the truly interesting anomalies: the rare, new, or unexpected objects that could rewrite our understanding of the cosmos. This is where the work of Michelle Lochner and her colleagues comes in, offering a new way to bridge the gap between raw computing power and human curiosity.

The researchers developed a tool called astronomaly, a system designed to help scientists sift through these massive datasets efficiently. Instead of trying to teach the computer exactly what a new type of object looks like—which is impossible if no one has ever seen one before—the system uses a method called active learning. In this process, the computer scans the data and presents a small selection of the most promising candidates to a human scientist. The scientist then labels these as interesting or uninteresting. The computer learns from these choices, refining its understanding of what the human finds valuable, and then sorts the rest of the data accordingly. This creates a partnership where the machine handles the heavy lifting of sorting, while the human provides the intuition to spot the truly significant discoveries.

This approach has already led to real findings. Using astronomaly, the team identified a new type of radio source named SAURON, which stands for a Steep and Uneven Ring of Nonthermal radiation. This object, found in data from the MeerKAT Galaxy Cluster Legacy Survey, does not look like any known source and might be the remnant of two supermassive black holes merging. The system also helped uncover unusual variable stars and galaxy mergers that had been missed by traditional methods. In one instance, the tool analyzed nearly 4 million optical images from the Dark Energy Camera Legacy Survey, highlighting a dozen specific sources that human experts found most intriguing, including objects that defied standard classification.

A key part of this success lies in how the computer understands the data. For decades, scientists had to manually design the features the computer should look for, such as the shape of a galaxy or the brightness of a star over time. This was a difficult step that often limited what the computer could find. The new work shows that using deep learning, a more advanced form of machine learning, allows the computer to figure out these features on its own. By training a computer to recognize complex patterns in images, the system can create a rich, internal map of the data. When the researchers applied this method, they found that the computer could group similar galaxies together and separate the unusual ones with far greater accuracy than before.

However, the researchers discovered a subtle but important twist in how these advanced systems work. When using older, manually designed features, strange objects tended to sit on the very edge of the data map, making them easy to spot as outliers. But with the new deep learning methods, the strange objects were often found deep within the map, mixed in with the normal ones. A computer looking only for things that sit on the edge would miss them. To solve this, the team refined their approach with a version of the tool called astronomaly: protege. Instead of just looking for statistical oddities, this version focuses entirely on what the human user finds interesting, learning directly from their feedback to navigate the complex map of data.

The paper concludes that the future of discovery in astronomy depends on this specific type of partnership between human and machine. As the flood of data from new telescopes grows, the ability to automate the search for the extraordinary will be essential. The researchers argue that because "interesting" is a subjective quality that varies from scientist to scientist, the best path forward is not to try to define it perfectly in code, but to build systems that learn it from us. By combining the speed of machines with the intuition of human experts, astronomers can ensure that the next great discovery, perhaps as surprising as the first pulsar found by Jocelyn Bell Burnell in 1967, is not lost in the noise.