Imagine staring into the vast cosmic ocean, where NASA's TESS telescope has been quietly scanning millions of stars, capturing fleeting shadows that hint at worlds beyond our imagination. For years, astronomers sifted through this treasure trove manually, spotting thousands of potential exoplanets—but many gems slipped through the cracks, buried under noise, false signals, and sheer data volume. Enter RAVEN AI, a groundbreaking pipeline that's just changed the game. Published today in Monthly Notices of the Royal Astronomical Society (MNRAS), researchers at the University of Warwick unleashed RAVEN on TESS's first four years of data, unearthing and validating 118 new exoplanets[1]—including 31 entirely new detections—plus over 2,000 high-quality candidates, nearly 1,000 previously unknown.[3] This isn't just a discovery; it's RAVEN AI exoplanets proving AI can outpace humans in decoding the universe's biggest datasets. Buckle up—let's dive in.
What is RAVEN? The AI Pipeline Revolutionizing Exoplanet Hunting
RAVEN stands for RAnking and Validation of ExoplaNets, a fully automated, end-to-end AI system custom-built for NASA's Transiting Exoplanet Survey Satellite (TESS) data.[4] Unlike traditional tools that handle one step—like detection or vetting—RAVEN does it all: spotting transit signals, classifying them as real planets or fakes, and statistically validating them with sky-high confidence.
Here's how it works, broken down simply:
- Detection: RAVEN scans light curves from TESS's Full Frame Images (FFIs) using a GPU-accelerated Box Least Squares (BLS) algorithm. It targets periods from 0.5 to 16 days—focusing on close-in worlds—and picks the top 5 signals per star, filtering out junk like harmonics or aliases.[1]
- Classification: Machine learning magic! Trained on hundreds of thousands of realistic simulations (planets plus 8 false positive scenarios like eclipsing binaries, background eclipses, and stellar noise), it uses Gradient Boosted Decision Trees (XGBoost) and Gaussian Process classifiers. These spit out probabilities for each scenario. Add priors (e.g., planet occurrence rates, centroid shifts), and you get a raven probability—the minimum posterior across false positives.[4]
- Validation: Candidates need raven probability ≥ 0.99 (and Planet-NSFP ≥ 0.99), planet radius < 8 Earth radii (R⊕), at least 3 full transits, and more vetting checks (e.g., depth variation <50%, S/N ≥3). Refined with juliet fits for precise parameters.
The results? 91% overall accuracy on 1,361 pre-classified TESS Objects of Interest (TOIs), with 97% precision at 0.9 threshold.[4] "RAVEN allows us to analyse enormous datasets consistently and objectively," says senior author Dr. David Armstrong. It's open-source and cloud-hosted—perfect for astronomers everywhere. See our guide on AI tools for astronomy to get started.
RAVEN processed 2.2 million main-sequence stars (Gaia-characterized, sectors 1-55), recovering 85% of TOIs and 54% of Community TOIs (CTOIs) in its period range.[1]
TESS Data: The Goldmine RAVEN Just Cracked Open
TESS, launched in 2018, scans the entire sky for transits—tiny dips (often <1%) in starlight as planets cross their host stars. Its FFIs cover 2.2 million stars here, but raw data is messy: stellar variability, instrument noise, cosmic rays.
Past pipelines like SPOC flagged ~10,000 candidates, but validation lagged—thousands remain unconfirmed. RAVEN changes that, delivering:
| Sample | Size | Stars | New? | Key Notes |
|---|---|---|---|---|
| BLS Candidates | 14,815 | - | - | Planet-NSFP ≥0.9[1] |
| Highly Ranked | 3,899 | - | - | raven ≥0.9 |
| Vetted | 2,170 | 2,112 | ~1,000 | High-probability, ≥2 transits; online tables[1] |
| Validated | 143 (118 new) | 140 | 31 | ≥0.99 prob, R_p <8 R⊕; phase-folded LCs available[1] |
| Large Radii (>8 R⊕) | 207 | - | 6 | Follow-up targets (e.g., Jupiters)[1] |
| Mono-Transits | 4 | 4 | All new | Single events, high sig; e.g., TIC 103629155[1] |
This uniform sample beats biased catalogs, enabling true demographics. Around 9-10% of Sun-like stars host a close-in planet.
Breakthrough Discoveries: Rare Worlds RAVEN Unearthed
RAVEN didn't just count planets—it spotlighted exotics:
-
Ultra-Short-Period (USP) Planets (<1 day orbit): Scorching super-Earths defying evaporation. Examples:
Planet Period (days) Radius (R⊕) Host Notes TOI-7008 b 0.84 ~4.95 Faint G=13.44 mag, Neptunian desert! [1] TOI-5736 b 0.65 ~1.52 Super-Earth[1] TIC 18942729 b 1.00 ~2.00 Multi-system[1] -
Neptunian Desert Planets: Neptune-sized (3-6 R⊕) worlds scarce at short periods (~2-4 days) due to photoevaporation/high-eccentricity migration. RAVEN validated several, including TOI-5486 b (P=2.02 d, 3.97 R⊕), TOI-4030 b (2.51 d, 5.85 R⊕). First direct measure: 0.08 ± 0.01% occurrence around Sun-like stars.[5] "For the first time, we can put a precise number on just how empty this ‘desert’ is,” says Dr. Kaiming Cui.
-
Multi-Planet Systems: New pairs like TOI-1839 (sub-Neptunes at 1.42 d / 4.02 d), TOI-4156 (4.46 d / 12.82 d).
These fill gaps in the radius-period plane (see Fig. 5 in the paper), probing formation theories.[1]
Why This Matters: AI Surpassing Humans in Astronomy
Astronomy generates petabytes yearly—TESS alone has FFIs from millions of stars. Humans vet ~hundreds/year; RAVEN scales to millions objectively. It recovers knowns flawlessly, flags false positives (e.g., only 1 FP in large-radius TOIs), and boosts TESS to Kepler-level demographics.
Implications:
- Planet Formation: USPs/Neptunian desert test evaporation/migration models.
- Future Missions: Preps targets for JWST, PLATO. Interactive catalogs released!
- AI Tools Boom: Try ExoMiner or AstroLLM for similar wins. Check our review of top AI astronomy software—includes affiliate picks like Python libs with GPU accel.
"Using our newly developed RAVEN pipeline, we were able to validate 118 new planets... This represents one of the best characterised samples," beams lead author Dr. Marina Lafarga Magro.
The Road Ahead: More Worlds with RAVEN and Beyond
RAVEN's just starting—future runs on full TESS (sectors 56+), short-cadence data, evolved stars. Demographics paper (DOI: 10.1093/mnras/stag022) dives deeper. Pair with radial velocities for masses; hunt Earth-like hints.
This is transformative: AI + massive data = exoplanet revolution. Tools like RAVEN make astronomy accessible—download it, run on your GPU.
FAQ
What exactly did RAVEN discover in TESS data?
RAVEN validated 118 new exoplanets (total 143, including 31 brand-new), 2,170 vetted candidates (~1,000 new), plus bonuses like 207 large-radius and 4 mono-transits. Focused on 0.5-16 day periods.[1]
### How rare are Neptunian desert planets according to RAVEN?
0.08% of Sun-like stars host them—first precise TESS measure, matching theory.
### Can I use RAVEN myself?
Yes! It's publicly released as a cloud app. Great for TESS FFIs; code on GitHub/arXiv. Needs Python, GPUs for speed.[4]
### Why focus on short-period planets?
They're brightest/easiest to detect, reveal hot Jupiters' origins, and fill demographics gaps. RAVEN nails 9-10% occurrence for close-ins.
What do you think—will AI like RAVEN find the first Earth 2.0? Drop your thoughts below!