A Teenager Just Unlocked a Decade of Astronomical Data No One Could Read
A 17-year-old from Pasadena High School has accomplished what professional astronomers couldn't: extracting 1.5 million previously hidden variable cosmic objects from the NEOWISE telescope's decade-long infrared survey. Working under mentor Davy Kirkpatrick at Caltech, Matteo Paz developed an artificial intelligence model called VARnet that cracked open an archive containing 200 billion individual detections—data so vast that conventional analysis methods simply couldn't handle it.
The breakthrough matters because variable objects—quasars that flicker, stars that pulse, binary systems that eclipse—hold keys to understanding stellar evolution, transient events, and the dynamic universe. Until now, these objects were buried in plain sight, their signal lost in the noise of sheer abundance. Paz's work transforms a storage problem into actionable science.
How a Summer Project Became Peer-Reviewed Research
Paz arrived at Kirkpatrick's lab in summer 2023 through Caltech's Summer Research Connection program, already an unusual student: he'd completed AP Calculus in eighth grade and was already studying undergraduate-level mathematics. His secondary school elective in coding and theoretical computer science had introduced him to machine learning. On day one, he told Kirkpatrick he wanted to publish a paper. Kirkpatrick didn't laugh.
Kirkpatrick's mentoring philosophy was shaped by his own experience—a Tennessee chemistry and physics teacher named Marilyn Morrison identified his potential and mapped his path to becoming an astronomer. Decades later, he channeled that same approach with Paz, connecting him to Caltech researchers Shoubaneh Hemmati, Daniel Masters, Ashish Mahabal, and Matthew Graham, who provided expertise in machine learning for time-domain astronomy.
The collaboration revealed a critical limitation: NEOWISE's scanning pattern, centered on great circles around the Sun, systematically misses objects that flash once and fade, or change gradually over years. No fully automated catalog based on NEOWISE alone will catch every variable phenomenon. That constraint shapes what Paz's 1.5 million candidates actually represent.
The Technical Accomplishment
VARnet processes astronomical time-series data through three integrated stages: wavelet decomposition reduces spurious measurements, a modified discrete Fourier transform extracts periodic features from irregularly sampled light curves, and a convolutional neural network classifies each source into four categories—non-variable, transient events like supernovae, intrinsic pulsators, or eclipsing binaries.
The model runs at less than 53 microseconds per source on a GPU with 22 gigabytes of VRAM, achieving an F1 score of 0.91 on validation data of known variables. That efficiency matters because it scales: Paz's approach can process the entire NEOWISE dataset, something that would have taken conventional methods years, if they could work at all. The Astronomical Journal published the complete technical specifications in peer review, legitimizing both the method and the teenager who built it.
What Comes Next
The 1.5 million flagged sources are candidates, not confirmations. Some will be known objects characterized in infrared wavelengths for the first time. Some will be false positives. Some fraction—the size of which remains unknown—will be genuine discoveries of quasars, variable stars, and transient events requiring follow-up observation. The full catalog is scheduled for release in 2025, providing the astronomical community a dataset large enough to support statistical studies of infrared variability across the entire sky rather than piecemeal analyses. Paz, now a Caltech employee at IPAC while finishing high school, has already begun discussing potential applications beyond astronomy—time-series analysis in financial markets, atmospheric pollution tracking, and other fields where periodic patterns matter. The teenager who walked into a lab with ambition just handed astronomers a decade's worth of new work.





