High Schooler's AI Model Uncovers 1.5 Million Hidden Cosmic Objects in NASA Archives

In early 2026, the Vera C. Rubin Observatory in the high deserts of Chile switched on its highly anticipated real-time alert system, sending astronomers worldwide an astonishing 800,000 notifications in a single night. These alerts flagged exploding stars, active black holes, and passing asteroids, marking a new era of high-speed astronomy. Yet, the foundation for this massive observational shift wasn't built entirely by a veteran astrophysics team working inside a government laboratory. Instead, a crucial piece of the underlying baseline data that makes these alerts possible was mapped by an 18-year-old high school student from California, proving that the tools of modern discovery are shifting.[1]

Matteo Paz, a senior from Pasadena High School, successfully identified 1.5 million previously unknown cosmic objects by applying a custom artificial intelligence model to archived space data. Mining nearly 200 terabytes of dense, noisy information from NASA's retired NEOWISE telescope, Paz achieved a scale of discovery that fundamentally expands the known catalog of the universe. His work bypassed the traditional decades-long slog of manual astronomical observation, utilizing advanced machine learning to spot patterns that human eyes and standard software had completely missed.[2][7]

The breakthrough has sent significant ripples through the global scientific community, validating the power of student-led research. It earned Paz the $250,000 first-place prize at the prestigious Regeneron Science Talent Search, resulted in a rare single-author publication in The Astronomical Journal, and recently drew formal public praise from NASA leadership. For a high schooler to not only participate in data analysis but to single-handedly architect a pipeline that uncovers over a million new celestial bodies is an unprecedented milestone that has forced major institutions to rethink how they engage with open data.[1][3][6]

The project originally began in the summer of 2022 when Paz joined the Planet Finder Academy at the California Institute of Technology. Under the direct mentorship of Davy Kirkpatrick, a senior scientist at Caltech's Infrared Processing and Analysis Center (IPAC), Paz was initially tasked with a relatively modest goal: looking for variable stars in a small, manageable slice of the night sky. The idea was to manually identify a few interesting objects to show the broader astronomical community the untapped potential hiding within the massive archives of retired space missions.[2][6]

Kirkpatrick desperately wanted to extract more value from the NEOWISE mission. Originally launched in 2009 to hunt for near-Earth asteroids and comets, the infrared telescope had scanned the entire sky for over a decade, amassing nearly 200 billion rows of precise thermal measurements. While tracking those local space rocks, the telescope inadvertently captured the faint, fluctuating heat signatures of distant quasars, supernovae, and eclipsing binary stars. Astronomers knew these variable objects were hidden in the data, but the sheer volume of the archive made extracting them seem like an impossible task.[4][6]

The core problem was one of overwhelming scale. The dataset was simply far too massive for manual human analysis, and traditional algorithmic software struggled to filter out the background noise to find the faint, irregular signals of these distant objects. Drawing on a remarkably strong background in theoretical math and coding, Paz decided to abandon the manual search entirely. Instead, he set out to build an automated machine-learning pipeline capable of processing the entire 200-billion-row archive without human intervention, turning a summer project into a massive computational challenge.[3][5]

The resulting artificial intelligence model, dubbed VARNet, represented a highly sophisticated approach to data mining. It combined wavelet decomposition to handle the dataset's inherent noise with a novel Fourier transform technique designed to extract key features related to light variability. Deep learning algorithms then processed these extracted features, allowing the system to achieve sub-millisecond processing speeds per light curve on standard graphics processing units. This meant the model could evaluate the entire history of a star's brightness in a fraction of a second, scaling up to handle the entire sky.[3][5]

The model's incredible computational efficiency allowed Paz to analyze the entire NEOWISE archive in a matter of weeks. VARNet ultimately flagged 1.9 million potential objects of interest exhibiting variable behavior. When this list was meticulously cross-referenced with existing astronomical databases, roughly 1.5 million of these sources showed absolutely no prior record. They were immediately distinguished as fresh targets for astrophysical research, instantly providing the scientific community with a massive new catalog of potential binary systems, pulsating stars, and active galactic nuclei.[3]

Released publicly to the scientific community in late 2025, this massive catalog is already driving active observations at major facilities around the globe. It provides an immediate, highly detailed roadmap for studying rare transients and cataclysmic variables—objects that change too subtly or unpredictably for standard sky surveys to catch. By establishing a baseline of what is normally fluctuating in the night sky, astronomers can more easily spot the truly anomalous events that signal major cosmic phenomena, saving thousands of hours of telescope time.[1]

The timing of the catalog's release is absolutely critical for the next generation of astronomy. When the Vera C. Rubin Observatory begins its massive ten-year Legacy Survey of Space and Time later this year, it is expected to issue up to seven million alerts every single night. To make sense of that unprecedented data deluge, astronomers desperately need the kind of baseline variable data that Paz's AI model has provided. Without it, the new observatories would be overwhelmed by false positives, flagging known variable stars as new discoveries.[1]

Beyond the immediate benefits to astrophysics, the underlying machine-learning algorithm has significant cross-disciplinary potential. Because VARNet is fundamentally designed to analyze noisy, time-based signals and detect subtle anomalies, researchers note it could eventually be adapted for a wide variety of terrestrial applications. Experts suggest the architecture could be utilized for environmental monitoring, high-frequency finance, or neuroscience, where tracking subtle temporal changes over massive datasets often reveals critical underlying patterns that traditional statistical methods miss. The ability to isolate a faint signal from an ocean of noise is a universal challenge in modern data science, making Paz's astronomical tool highly relevant to Earth-bound industries.[1][5]

Ultimately, the achievement serves as a powerful reminder of how technology is fundamentally reshaping the landscape of academic research. As open-source data proliferates and artificial intelligence capabilities expand, the barrier to entry for world-class discovery is rapidly lowering. It proves that revolutionary scientific leaps no longer require multi-billion-dollar institutional supercomputers or decades of tenure. Sometimes, they just require open data, accessible algorithms, and a fresh perspective from an independent researcher willing to ask new questions of old information.[1][5]