AI Uncovers 118 New Exoplanets with RAVEN, Achieving 91% Overall Accuracy

RAVEN: A New AI‑Driven Exoplanet Screening and Validation Pipeline

The authors present RAVEN (RAnking and Validation of ExoPlaNets), an end‑to‑end workflow for TESS candidates that replaces reliance on threshold‑crossing‑event (TCE) data with a large synthetic training set, dramatically expanding the parameter space of planetary and false‑positive (FP) scenarios covered by the machine‑learning models.

Dataset Construction

Input light curves are generated from TESS full‑frame images (FFI) using aperture photometry. Sectors 1‑27 have 30‑min cadence, 28‑55 have 10‑min cadence, and the second extended mission (from sector 56) provides 200‑s cadence; the study uses data up to sector 55.

Synthetic training data are created by injecting simulated events—planet transits, eclipsing binaries (EB), hierarchical eclipsing binaries (HEB), hierarchical transiting planets (HTP), background eclipsing binaries (BEB), and background transiting planets (BTP)—into SPOC FFI light curves. Stellar hosts are randomly drawn from the well‑characterised TESS Input Catalog (TIC), yielding 1,200,520 synthetic stars.

For nearby false positives (NFPs), three scenarios are modelled: nearby transiting planet (NTP), nearby eclipsing binary (NEB), and nearby hierarchical eclipsing binary (NHEB).

Machine‑Learning Models

RAVEN combines two classifiers:

Gradient‑Boosted Decision Tree (GBDT) : implemented with XGBoost, each new tree learns from the residuals of the previous ensemble, minimizing a predefined loss function.

Gaussian Process (GP) classifier : uses variational inducing points (Hensman et al.) to scale inference, mapping GP outputs to probabilities via a Bernoulli likelihood.

For each candidate, eight FP scenario posterior probabilities are computed; the minimum of these values defines the RAVEN probability, representing the lowest confidence that the candidate is a true planet.

Training, Calibration, and Early Stopping

Hyper‑parameter search is performed on the synthetic set, focusing on the three most common FP scenes (EB, NEB, NHEB) while keeping parameters consistent across scenes to avoid over‑fitting. Early stopping halts training when validation loss does not improve by at least 0.0001 for 20 consecutive iterations.

External Validation on TESS TOIs

The pipeline is evaluated on an independent set of 1,361 pre‑classified TOIs (705 known or confirmed planets, 630 FP, 26 FA). Performance metrics—accuracy, AUC, precision, recall—are reported in the table below.

Across all FP scenarios, AUC exceeds 97% (most >99%). Precision approaches 99% in every scene, reflecting the pipeline’s ability to reject false positives without sacrificing true‑planet recovery.

When applied to the full TOI sample, 93.8% of FP events receive a minimum posterior probability below 0.5 (69.7% below 0.01). The average FP probability is 0.076, median 0.00022. For FA TOIs, 23 of 26 have probabilities below 0.5, confirming RAVEN’s effectiveness at flagging spurious alerts.

Discovery of New Exoplanets

RAVEN identified 118 previously unknown exoplanets and generated a catalog of over 2,000 high‑quality planet candidates, nearly 1,000 of which had never been reported. After filtering, 705 planets (420 with probability >0.9, 210 with >0.99) remain, demonstrating the pipeline’s utility for large‑scale planet discovery.

Broader Implications for AI in Astronomy

The authors argue that AI is transitioning from a data‑processing aid to a foundational infrastructure that reshapes scientific discovery. They contrast traditional physics‑based analyses with modern multimodal AI models, noting limitations of current contrastive approaches and highlighting emerging foundation models such as AION‑1 that integrate images, spectra, and catalog data.

References: RAVEN paper (arXiv:2509.17645), AION‑1 foundation model (OpenReview), and related works on AI‑driven strong‑lens quasar identification (arXiv:2511.02009).