Imagine uncovering 1.5 million hidden secrets in the cosmos—a feat so extraordinary that it left astronomers in awe. That’s exactly what happened when a 17-year-old American high school student, Matteo Paz, teamed up with Caltech scientist Davy Kirkpatrick to revolutionize our understanding of the universe. But here’s where it gets even more astonishing: these secrets had been sitting in plain sight for over a decade, buried within the vast archives of the NEOWISE telescope’s infrared data. How did they do it? And why did it take a teenager to crack the code? Let’s dive in.
The story begins with the NEOWISE telescope, which spent 10.5 years scanning the entire sky, collecting a staggering 200 billion individual detections. This data, primarily used for asteroid tracking, was a treasure trove of untapped potential. Among the billions of observations were variable objects—quasars flickering, stars pulsing, and binaries dimming as they eclipsed—all hidden in the noise of their sheer abundance. For years, Kirkpatrick had wondered what other discoveries lay dormant in this dataset. His goal? To create a comprehensive catalog of every infrared source that changed brightness over time.
And this is the part most people miss: the dataset had grown so massive that traditional methods were no longer feasible. Kirkpatrick recalls, ‘We were approaching 200 billion rows of data—every single detection over a decade.’ His initial plan was modest: analyze a small patch of sky by hand to prove the concept. But then Paz walked into his lab and proposed something far more ambitious.
Paz, a mathematical prodigy who had completed AP Calculus by eighth grade, was no stranger to Caltech. He had attended public stargazing lectures since childhood and was already studying undergraduate-level mathematics. An elective course in coding and theoretical computer science introduced him to machine learning, setting the stage for his groundbreaking work. On their first day together, Paz told Kirkpatrick he wanted to publish a paper. Kirkpatrick, impressed by Paz’s drive, gave him the freedom to explore. ‘He allowed an unbridled learning experience,’ Paz said. ‘That’s why I’ve grown so much as a scientist.’
The result? VARnet, a machine learning model that processes astronomical time series data in just 53 microseconds per star. VARnet works in three stages: wavelet decomposition filters out noise, a modified discrete Fourier transform identifies periodic patterns, and convolutional neural networks classify each source into one of four categories—non-variable, transient events like supernovae, intrinsic pulsators, or eclipsing binaries. Published in The Astronomical Journal, VARnet achieved an F1 score of 0.91 on known variable objects, proving its scalability to the entire NEOWISE dataset.
But here’s the controversial part: while VARnet flagged 1.5 million potential variable objects, not all are confirmed discoveries. Some are known objects seen in infrared for the first time, others are false positives, and a fraction are genuinely new detections. This has sparked debate: Is the model overestimating, or are we underestimating the universe’s complexity? The full catalog, set for release in 2025, will provide astronomers with the largest dataset ever for studying infrared variability across the sky.
Kirkpatrick’s mentorship, inspired by his own high school teacher in Tennessee, played a pivotal role. ‘I wanted to pass on that same kind of mentoring,’ he said. By connecting Paz with Caltech experts in machine learning and astronomy, Kirkpatrick ensured the project’s success. Yet, they discovered a limitation: NEOWISE’s observational rhythm couldn’t detect objects that flash once or change gradually over years. Some variable phenomena remain invisible, leaving room for future exploration.
Paz, now a Caltech employee while finishing high school, sees applications beyond astronomy. ‘VARnet can be used for any time-domain study,’ he said. ‘From chart analysis to atmospheric effects like pollution, where periodic patterns are key.’ This raises a thought-provoking question: Could AI models like VARnet transform fields beyond science, reshaping how we analyze temporal data?
What do you think? Is this the future of discovery, or are we relying too heavily on AI? Share your thoughts in the comments—let’s spark a conversation about the boundaries of human and machine collaboration in uncovering the universe’s secrets.
