Matteo Paz (left) with Caltech President Thomas F. Rosenbaum | Photo: Society for Science via Caltech
Matteo Paz, a Pasadena High School senior whose interest in astronomy was sparked by Caltech public lectures, used AI to develop a computational model while participating in Caltech’s Summer Research Connection outreach program. Under the mentorship of IPAC astronomer Davy Kirkpatrick, Paz analyzed over a decade’s worth of NASA’s NEOWISE infrared survey data. While the telescope primarily sought asteroids, its dataset of nearly 200 billion measurements offered vast potential for data mining variable objects, far exceeding manual analysis capabilities. Paz’s model identified roughly 1.5 million celestial objects previously unknown to scientists, employing advanced Fourier and Wavelet transforms (leveraging a specialized PUSD math background) to extract faint signals from noisy data. This achievement, detailed in his single-author publication in The Astronomical Journal, also secured Paz the top $250,000 prize in the 2025 Regeneron Science Talent Search. Now an employee at Caltech/IPAC, Paz plans to publish the full object catalog with Kirkpatrick in 2025.
Dealing with scale
The core challenge lay in the sheer scale of the NEOWISE dataset. Its characteristics make it far too unwieldy for traditional analysis methods. Paz’s AI model, dubbed VARnet, was specifically developed for this scale. It combines wavelet decomposition to handle noise with a novel Fourier transform technique (FEFT) for extracting key features related to variability. Deep learning then processes these features, allowing VARnet to achieve sub-millisecond processing per light curve on GPUs.
Paz analyzed the 1.9 million total sources the model flagged. About 1.5 million of these showed no prior record in existing catalogs, immediately distinguishing them as fresh targets for research. This massive influx of newly identified variable objects spans diverse astrophysical phenomena, from potential supernovae to slowly pulsating stars with decade-long cycles. Preliminary classification suggested at least four broad categories, hinting at potentially up to ten distinct classes once follow-up observations confirm their nature.
Scientific Impact and Recognition
This breadth of discovery underscores the value of exploiting the time-domain dimension of large astronomical archives. VARnet effectively transformed a historical dataset into a trove of valuable data, creating the new VarWISE catalog. Researchers worldwide can now reference this catalog to select compelling objects for targeted studies of stellar evolution, black hole accretion, and other dynamic cosmic events using ground-based or space-based missions.
Beyond the discoveries themselves, Paz’s single-author publication in The Astronomical Journal demonstrates the scientific rigor behind VARnet. Passing formal peer review as a high school student highlighted the robustness of his algorithms and validation process. His subsequent first-place finish in the 2025 Regeneron Science Talent Search brought significant public attention, emphasizing the achievement’s rarity and the pivotal role of Caltech’s mentorship programs.
Paz plans to release the full VarWISE catalog to the community, complete with updated classifications and cross-matches. By making these results publicly available, he aims to spur follow-up campaigns that might confirm the specific astrophysical nature of each candidate—especially those displaying unusual infrared variability. This approach showcases how advanced AI pipelines, tailored to unique scientific datasets, can trigger a wave of new discoveries and collaborations across the astronomical community.
