A California high school student has achieved what seasoned astronomers spend careers pursuing: uncovering more than 1.5 million previously unknown celestial objects using an AI model he developed himself. His work, drawn from dormant NASA data, is already reshaping how scientists explore the infrared universe.
Paz didn’t stumble across his find by luck or accident. Instead, he engineered a bespoke machine learning pipeline to scan over a decade’s worth of NEOWISE infrared sky data—more than 200 billion rows of observations—sifting for subtle signals that had long gone unnoticed. Among them: flickers, pulses, and dimmings in the infrared sky that may point to quasars, supernovae, eclipsing binaries, and other cosmic variables that are as difficult to detect as they are scientifically important.
NASA’s NEOWISE mission, originally launched in 2009 to detect near-Earth asteroids, accumulated a vast repository of sky data before it was decommissioned. Yet only a fraction of that archive had been closely examined for time-domain phenomena—astronomical events that change brightness over time. NEOWISE was excellent at collecting data, but like many space missions, it left the challenge of interpretation to the future.
That future arrived, unexpectedly, through a public school student with a passion for astronomy and a knack for coding. Paz, now 17, had enrolled in Caltech’s Planet Finder Academy, a mentorship initiative designed to bring young minds into the orbit of cutting-edge astrophysics. His mentor, Davy Kirkpatrick, a senior scientist at Caltech’s Infrared Processing and Analysis Center (IPAC), had initially imagined they might hand-analyze a small subset of the data. Paz suggested building an algorithm to process all of it.
The result: an AI model that blends Fourier transform techniques with wavelet analysis, enabling it to detect fluctuations in the brightness of distant cosmic objects—variations that often hint at previously unseen astrophysical activity. Kirkpatrick admits he was astonished. “Matteo’s model began returning intriguing results almost immediately,” he said. “And it just kept getting better.
Scanning the Sky, One Pulse at a Time
Paz’s model is not just sophisticated—it’s versatile. Designed to pick up changes in light that occur over hours, weeks, or even years, the algorithm flagged millions of light curves that had been missed by human analysts and traditional pipelines. Some of the signals were fleeting, others were so gradual they escaped previous detection altogether.
These signals aren’t just cosmic curiosities. They may help astronomers refine how we understand stellar evolution, map galactic structures, or even identify rare events like gravitational lensing. Some may also represent entirely unknown classes of phenomena, waiting for further observation through next-generation telescopes like the Vera Rubin Observatory or NASA’s James Webb Space Telescope.
The full catalog of 1.5 million objects is expected to be publicly released in 2025. For now, the academic community is reviewing early results, with interest growing fast. Paz’s original paper, published in The Astronomical Journal, has sparked discussions among professional astronomers about the untapped potential of legacy data—especially when paired with fresh algorithms and outsider perspectives.
A Pipeline Built to Travel
What makes Paz’s discovery even more compelling is its transdisciplinary potential. His algorithm, designed for variable star detection, works on any kind of time-series data. That means its applications stretch far beyond space.
“Once you realize this is just about detecting change over time, the possibilities open up,” Paz said in a recent Caltech interview. “You could use it to monitor pollution patterns, analyze seismic data, or even track market volatility.”
Indeed, time-domain analytics—especially when built on clean, structured datasets like NEOWISE—are a growing trend across the sciences. NASA, for its part, has already begun investing more heavily in AI-enhanced astronomical workflows, as seen in its partnerships with researchers at IPAC and beyond. Paz’s work arrives as a clear proof-of-concept: trained properly, machine learning can do more than accelerate data analysis—it can transform it.
Mentorship, Math, and the Making of a Scientist
At the center of this story lies not just innovation, but mentorship. Kirkpatrick, known for his work on brown dwarfs and low-mass stars, has long advocated for bringing underrepresented voices into astronomy. “We have all this data and not enough people looking at it,” he said. “When I see talent like Matteo’s, I want to make sure it gets the support it deserves.”
Paz came prepared. He’s part of Pasadena Unified School District’s Math Academy, a public school program that accelerates mathematically gifted students through college-level material before high school graduation. By eighth grade, he had already completed AP Calculus BC. By senior year, he had published solo in a major scientific journal.