A groundbreaking application of machine learning has revolutionized the search for exoplanets, uncovering a record-breaking haul of over 10,000 transiting exoplanet candidates hidden in data from NASA's Transiting Exoplanet Survey Satellite (TESS). This discovery, led by researchers primarily from Princeton University, more than doubles the number of known TESS candidates and opens new frontiers in understanding planetary systems around faint stars.

The T16 Planet Hunt project analyzed an astonishing 83.7 million light curves from TESS's first cycle of observations, targeting stars as faint as 16th magnitude in the TESS band—far dimmer than typical searches focus on. Traditional methods prioritized brighter stars for easier detection and follow-up, leaving a vast reservoir of potential worlds undetected until now.

🌌 Decoding Transiting Exoplanets with TESS

Transiting exoplanets are worlds that pass in front of their host stars from our line of sight, causing periodic dips in the star's brightness. These dips, captured as light curves, provide key data on planet size, orbit, and stellar properties. NASA's TESS, launched in 2018, scans nearly the entire sky in 27-day sectors using four wide-field cameras, producing full-frame images (FFIs) every 30 minutes for bright stars and 10 minutes for fainter ones in later extended missions.

TESS has already confirmed 882 exoplanets, contributing 14% to the over 6,000 known as of late 2025. However, its FFIs hold untapped potential for faint-star transits, where planet occurrence rates suggest abundant short-period worlds. The mission's primary goal is finding small, Earth-sized planets around nearby bright stars for James Webb Space Telescope (JWST) follow-up, but secondary science like this planet hunt expands its legacy.

The Machine Learning Pipeline Behind the Discovery

At the heart of the T16 project is a semi-automated pipeline leveraging random forest classifiers (RFCs)—ensemble machine learning models that excel at handling noisy, high-dimensional data. First, light curves were uniformly detrended using the T16 dataset, correcting systematics across 54 million stars.

Transit searches employed CETRA for initial detection and Box-Least Squares (BLS) periodograms for parameters. Fifty-three features, including signal-to-pink-noise ratio, transit depth, and Gaia DR3 stellar properties like effective temperature (Teff), fed into two RFCs: a base model for all magnitudes and a faint-star specialist for T>14.5 mag. A secondary RFC vetted candidates, flagging non-transits with high recall.

Contamination checks used graphs of nearby sources, BATMAN models ensured physical radii (<2.5 RJup), and manual inspection of vetting plots (phase-folded curves, residuals, spectra) ensured quality. This hybrid human-AI approach yielded 11,554 candidates: 10,091 new, 1,052 known, and 411 single-transits (0.5-27 day periods).

Princeton University: Hub of Exoplanet Innovation

Graduate student Joshua T. Roth led this effort from Princeton University's Department of Astrophysical Sciences, collaborating with professors Joel D. Hartman and Gáspár Á. Bakos. Princeton's storied astrophysics program, home to pioneers in exoplanet detection, provided the computational resources and expertise for processing massive datasets.

Co-authors hail from UCLA, Carnegie Science, MIT's Kavli Institute, and Universidad de Concepción, showcasing interdisciplinary academia. Carnegie Observatories handled radial-velocity confirmation using Magellan telescopes. Such collaborations highlight how university-led projects drive NASA missions forward, training next-gen astronomers in ML and big data.

This work exemplifies higher education's role in space science, where grad students like Roth publish in top journals like The Astrophysical Journal Supplement Series, boosting careers in academia and observatories.

Photo by Joshua Hoehne on Unsplash

Profiles of the Newly Discovered Candidates

These candidates orbit faint stars (mostly T>13.5 mag), with 97.7% gas giants, 1.5% Neptunes, 0.64% sub-Neptunes, and 11 super-Earths. Notably, 737 around low-metallicity ([Fe/H]<-1) hosts and 66 ultra-short-period (<1 day). Orbital periods cluster short, implying hot environments unlikely for life, but valuable for demographics.

Single-transits hint at longer periods, resolvable with TESS Year 2 data. The catalog, public as CTOIs with Gaia/TIC IDs, invites community validation.

• Planetary radii mostly <2 RJup, vetted against blends.
• 87% multi-transit detections, robust periods.
• Faint-star bias reveals occurrence rates missed before.

Validation Milestone: The Hot Jupiter TIC 183374187 b

Pipeline prowess proven by confirming TIC 183374187 b: a hot Jupiter (1.252 RJup, 0.56 MJup, 5.059-day orbit) around a metal-poor thick-disk star 3,950 light-years away. Magellan/PFS radial velocities ruled out blends, matching transit ephemerides.

This first validation bodes well for others, though full confirmation demands years of spectroscopy. Read the full paper for details: T16 Planet Hunt arXiv preprint.

Reshaping Exoplanet Demographics and Astrophysics

By targeting faint stars, T16 expands the census, enabling precise occurrence rates. Early hints: higher giant planets around dim hosts, informing formation theories. Low-metallicity systems probe early universe planets.

Complements Warwick's RAVEN AI, validating 118 TESS planets including Neptunian-desert fillers. Doubles TESS candidates, priming JWST/PLATO targets.

Challenges in Confirmation and False Positives

Not all candidates are planets; eclipsing binaries (EBs), blends loom. Vetting caught many, but manual limits (4,500 inspected) suggest ~10-20% false positives. Tools like TRICERATOPS aid, but hot Jupiters tricky.

Future: Year 2 analysis could double yield with better S/N. Ground follow-up strained by faintness, needing ELTs.

Photo by Mauro Romero on Unsplash

Boosting Careers in Astronomy and Data Science

This discovery underscores demand for ML-savvy astronomers. Universities like Princeton offer PhDs blending astrophysics/compsci, leading to postdocs at Carnegie/MIT. Explore NASA Exoplanet Archive for data skills.

Higher ed implications: Funds ML tools, interdisciplinary programs. Grads enter tenure-track, observatories, tech (AI for space).

Future Horizons: PLATO, JWST, and Beyond

T16 primes ESA's PLATO (2026+), surveying brighter fields longer. JWST atmospheres on confirmed candidates reveal compositions. ML evolves: CNNs, transformers next.

Universe teems; academia drives discovery.