science 5 min read

A Teenager's AI Just Proved Citizen Science Is Real

Matteo Paz built a machine-learning pipeline in six weeks that found 1.5 million flickering objects in NASA's archived sky data—now professional astronomers are using his catalog and asking how many more untapped datasets exist.

  • Astronomy
  • NASA
  • Artificial Intelligence
  • Citizen Science
  • Data Science

The pipeline was built in six weeks

Matteo Paz was fifteen when he joined the Planet Finder Academy, a Caltech-affiliated program that pairs students with working astronomers. His mentor was Davy Kirkpatrick at the Infrared Processing and Analysis Center. The assignment: sift through roughly two hundred billion rows of measurements from NASA’s NEOWISE telescope, a satellite launched in 2009 to track near-Earth asteroids but which quietly accumulated a decade of full-sky infrared observations.

Paz did not take the intended route. Instead of studying a small subset, he spent six weeks building a machine-learning pipeline that scanned the entire archive for faint, variable light sources — objects whose brightness flickers, pulses, or fades in ways that standard detection software overlooks. The system flagged 1.5 million of them. The catalog, released publicly in late 2025, is now driving active observations at major facilities including the Vera C. Rubin Observatory.

The result matters less for the number itself than for what it proves: the bottleneck in modern astronomy is no longer computational power or access to data. It is time. Institutions sit on petabytes of archived sky surveys because they lack the personnel to process them. A teenager with a laptop, open-source mathematical tools, and a structured mentoring relationship just showed that the gap can close fast.

How it actually works

Paz’s model relied on Fourier transforms and wavelet analysis — techniques for extracting periodic and transient signals from noisy time-series data. These are graduate-level tools, standard in signal processing but uncommon in introductory astronomy curricula. Paz learned them through the Pasadena Unified School District’s Math Academy, a public program for mathematically gifted students.

NEOWISE samples the sky in infrared at regular intervals. Most celestial objects appear steady. But quasars flicker as supermassive black holes consume matter. Binary stars eclipse each other. Supernovae flare and fade. Standard pipelines look for bright, predictable variations. Fainter or rarer signals slip through the cracks.

Paz’s algorithm did the opposite of chasing brightness. It looked for the subtle deviations — the flicker, the pulse, the slow dimming — across the full ten-year dataset. The result was a catalog of 1.5 million variable sources that had not been systematically identified before. Several Caltech researchers, including Shoubaneh Hemmati, Daniel Masters, Ashish Mahabal, and Matthew Graham, helped refine the system to handle the entire sky.

What happens next for the data

The Vera C. Rubin Observatory has already begun drawing on Paz’s catalog. On February 24, 2026, its real-time alert system fired roughly 800,000 warnings in a single night — exploding stars, variable objects, active black holes, asteroids. That was before Rubin’s main ten-year Legacy Survey of Space and Time began later that year. Once fully operational, Rubin is expected to generate up to seven million alerts every night. No human team can process that volume. Machine-learning brokers have to decide, in seconds, which signals deserve follow-up observation.

In other words, the approach Paz prototyped on archived data has quietly become the only way modern astronomy operates at scale. The telescopes have grown so powerful that artificial intelligence is no longer a convenience. It is the entry requirement.

NASA leadership has publicly acknowledged Paz’s contribution, describing it as evidence that major space institutions now recognize revolutionary discoveries can emerge from non-traditional sources. That acknowledgment carries weight. It signals a shift in how funding bodies and agencies evaluate where talent and insight come from.

The broader implication: what else is sitting untapped?

The immediate story is inspiring. The structural story is bigger. Nebulosity lies in archived datasets that professionals simply do not have time to examine. The Gaia mission, which has mapped over a billion stars, has released multiple data releases containing time-series photometry that remains largely unmined for transient and variable signals. Zwicky Transient Facility archives, Pan-STARRS sky surveys, and the archived TESS light curves all contain similar troves.

Each of these datasets shares the same problem NEOWISE faced before Paz: too much data, not enough hands. The difference now is that the barrier to building a mining pipeline is dramatically lower. Open-source libraries for time-series analysis, Fourier transforms, and wavelet decomposition are freely available. A motivated student with basic coding skills can prototype what once required a research group and grant funding.

Several questions remain unresolved. How representative is Paz’s 1.5-million-object catalog of the true population of variable sources in the infrared sky? Some detections may be instrumental artifacts or known object classes misclassified by the algorithm. Professional astronomers will need to validate a sample before the catalog enters textbooks. And Paz himself has said he plans to extend the approach beyond astronomy — to finance, environmental monitoring, neuroscience — areas where the same challenge of sifting through noisy time-series data applies.

What is clear is that the model is reproducible. The Planet Finder Academy pairing of students with institutional mentors provided the scaffolding. The open data provided the material. The AI pipeline provided the engine. Remove any one component and the result changes significantly. Keep all three and you get a high school student outperforming the manual processing plans of a professional research team.

Who wins and who loses

Students in under-resourced districts win. Public school programs that invest in advanced mathematics and computational training — like Pasadena’s Math Academy — gain a visible return on investment. Institutions that open their data, such as IPAC and NASA, gain unexpected analytical capacity without additional budget lines.

Traditional gatekeeping models lose. The assumption that serious astronomical discovery requires a PhD, a lab position, and institutional computing resources is no longer defensible. It never really was, but Paz’s work makes the contradiction impossible to ignore.

The next decade of sky surveys will amplify this pattern. Rubin alone will produce more data in a single observing season than humanity has collectively processed from optical and infrared surveys over the past fifty years. Human analysts will remain essential for design, validation, and interpretation. But the initial sifting — the detection of which flickering pixels merit a second look — belongs to algorithms. The question is no longer whether amateurs and students can build those algorithms. It is whether institutions will create pathways wide enough for them to matter at scale.