High school researcher Matteo Paz achieved a remarkable breakthrough in astronomy by developing an artificial intelligence algorithm capable of uncovering 1.5 million previously unknown celestial objects hidden within NASA’s NEOWISE infrared survey. While NEOWISE had spent more than a decade scanning the entire sky and collecting nearly 200 billion individual detections, most of its data on variable phenomena—objects that brighten, dim, pulse, or flare—remained unexamined. Paz recognized that this immense dataset, though unwieldy for manual inspection, was perfectly suited for machine learning. His resulting algorithm not only identified new astronomical sources but also expanded the scientific potential of the NEOWISE mission itself, ultimately earning him first place and a $250,000 award in the 2025 Regeneron Science Talent Search.
The foundation of Paz’s accomplishment comes from his ability to design a model that could process massive time-series datasets. NEOWISE repeatedly observed the same regions of the sky, capturing infrared brightness measurements that formed temporal “signatures” for millions of objects. Paz realized that these temporal patterns were ideal for differentiating variable stars, quasars, and other astrophysical phenomena. His algorithm began by ingesting raw NEOWISE detections and grouping them by sky coordinates, reconstructing each object’s full brightness history. This reconstruction step was essential because it converted billions of individual detections into coherent timelines that could reveal which celestial objects varied in measurable ways.
Once reconstructed, the algorithm performed an extensive cleaning process. Paz applied mathematical filters to eliminate noise from telescope jitter, sensor drift, and atmospheric interference. He also used statistical normalization to ensure that measurements from different years and observing conditions were comparable. With the dataset cleaned and stabilized, the algorithm next searched for key markers of variability: periodic oscillations, sudden spikes, long smooth fades, or eclipse-like dips. These patterns—often invisible to the human eye within such a large dataset—provided critical clues about the underlying astrophysical processes driving each object’s behavior.
To classify the detected signals, Paz incorporated machine-learning classifiers trained on known examples of variable stars and quasars. These classifiers compared new patterns against established categories, enabling the system to identify common types of variable objects while also flagging anomalies that might represent previously unstudied or rare forms of celestial objects. A major challenge was NEOWISE’s irregular observational cadence; the telescope revisited parts of the sky at uneven intervals, making traditional detection techniques unreliable. Paz overcame this by designing the algorithm to learn directly from the data rather than depending on fixed assumptions about timing, allowing it to recognize both rapid flashes and multi-year changes.
After two years of refinement, Paz scaled the algorithm to the full dataset and analyzed the output. The final system flagged 1.5 million candidates—each representing potential additions to a forthcoming catalog of variable NEOWISE sources. Beyond astronomy, Paz notes that the same approach can be applied to any field involving time-series data, such as stock-market fluctuations or atmospheric pollution cycles.
Through this achievement, Paz not only revealed a massive population of previously hidden celestial objects but also demonstrated the transformative power of AI in unlocking scientific insights buried inside vast archival datasets.
https://scitechdaily.com/teen-wins-250k-for-using-ai-to-discover-1-5-million-hidden-objects-in-space