A talented high schooler from Pasadena engineered an AI tool designed to scan one of the largest astronomical archives, identifying and categorizing 1.5 million possible variable celestial objects documented by NASA’s retired NEOWISE mission.
During his time at Pasadena High School, Matteo Paz created this innovative system while conducting research at Caltech. What began as a summer project over six weeks evolved into a solo-authored, peer-reviewed publication in The Astronomical Journal and secured Paz the $250,000 grand prize at the 2025 Regeneron Science Talent Search.
The figure of 1.5 million candidates demands clarification. These are potential variable sources flagged by Paz’s algorithms—many still require further study. A February 2026 report notes these flagged points need additional confirmation and classification, with some expected to be known objects or false positives.
The investigation originated from a simple question: what secrets lie hidden within over ten years of infrared data?
From a Six-Week Project to Managing Billions of Data Points
In 2023, Paz joined Caltech’s Summer Research Connection, collaborating with Davy Kirkpatrick, a scientist at IPAC with a deep interest in uncovering hidden information in the NEOWISE datasets. Kirkpatrick’s profile highlights his focus on brown dwarfs, stellar classification, and low-mass stars near the Sun.
NEOWISE operated for over a decade, mapping the entire sky in infrared light to spot asteroids and near-Earth objects, but its repeated scans also captured varying infrared emissions from distant phenomena such as quasars, supernovae, and eclipsing binary stars.
The sheer amount of data posed a major challenge.
By the time Kirkpatrick discussed this project, the NEOWISE database had grown to nearly 200 billion entries. Initially, the idea was for a student to manually examine a small sky region for variable stars and showcase findings for astronomers to explore further.
However, Paz envisioned a much broader approach.
Instead of limiting analysis to one section, he aimed to create an astronomical machine-learning system capable of automatic and expansive dataset processing. Caltech noted Paz’s advanced math background, having completed AP Calculus BC in eighth grade followed by studies in coding, theoretical computer science, and AI through his school’s Math Academy.
That summer, Paz began crafting the AI model while Kirkpatrick introduced him to the astronomical concepts behind the data. Additional mentorship came from Caltech astronomers Shoubaneh Hemmati, Daniel Masters, Ashish Mahabal, and Matthew Graham, all specialists in machine learning and variable astronomical phenomena.
“I’m incredibly grateful to have worked with Davy,” Paz shared with Caltech. Their collaboration extended beyond the initial program and into 2024, during which Paz mentored fellow students.
VARnet: Ultra-Fast Analysis of Astronomical Time-Series
Paz’s system, named VARnet, was engineered to rapidly examine time-series astronomical data.
It integrates wavelet decomposition, Fourier transforms, and neural network classifiers. The peer-reviewed article in The Astronomical Journal outlines VARnet’s ability to identify variable candidates from the NEOWISE Single-Exposure archive, with Paz credited as the sole author.
VARnet analyzes astronomical time-series where measurements occur at different intervals, classifying sources as non-variable, transient, pulsating, or eclipsing binaries.
On the specified hardware, VARnet classified each source in under 53 microseconds using a graphics card equipped with 22 GB of VRAM. It reached an F1 score of 0.91 on a test set containing known variable stars.
This remarkable speed enabled Paz to analyze far more than just a limited sky region. He fine-tuned the tool to process raw NASA NEOWISE data and detect subtle infrared brightness changes. By April 2025, Caltech revealed that VARnet had identified and sorted 1.5 million promising variable candidates.
While NEOWISE’s mission focused on nearby asteroids, its repeated sky scans also encompassed brightness variations from faraway sources. VARnet’s speed made it possible to generate an extensive catalog of possible variable objects, a task unfeasible with manual searching.
Potential Discoveries Await Confirmation
Descriptions of Paz’s findings vary across sources. Caltech initially called them 1.5 million newly uncovered objects but later clarified the algorithms had merely flagged and sorted 1.5 million potential variable objects.
These candidates require additional observations and classification before being formally recognized. Some may correspond to previously cataloged objects, others might be errors or false alarms.
VARnet’s capability is also limited by the NEOWISE observation schedule. The telescope’s observation cadence means it cannot effectively categorize objects that exhibit single, brief flashes or those changing very slowly.
If a source’s variability isn’t sampled frequently enough, VARnet won’t have sufficient data to identify its pattern.
Achievement Recognized With a Prestigious Prize
Paz’s relationship with Caltech dates back to childhood, attending public astronomy lectures with his mother. In 2022, he participated in Caltech’s Planet Finder Academy to deepen his understanding of astronomy and computer science.
The following year, he joined the Summer Research Connection, partnering with Kirkpatrick. This six-week endeavor culminated in a published paper showcasing the power of AI for celestial investigations.
His groundbreaking work earned Paz first place in the Regeneron Science Talent Search in March 2025, accompanied by a $250,000 prize. Caltech documented the award ceremony held on March 11, 2025.
By the following month, Paz was finishing his high school education while working at IPAC, a Caltech facility responsible for managing and analyzing data from NEOWISE and other space missions sponsored by NASA and the NSF. Caltech described this role as Paz’s first paid position.
What started as a project focused on a small area of the sky blossomed into an award-winning AI tool, a scientific publication, and a trove of 1.5 million AI-detected variable star candidates spanning over a decade of NEOWISE data.
- Categories:
- Science
0 comments
Sign in to Comment