Automatic search for transiting planets in TESS-SPOC FFIs with RAVEN: over 100 newly validated planets and over 2000 vetted candidates

Abstract: Space-based missions such as TESS are identifying a wealth of short-period ($\lesssim30$ d) transiting planets. Despite the growing number of confirmed and candidate planets, the sample is still incomplete and highly biased, challenging demographic studies. Moreover, there are still a large number of unconfirmed candidates that can end up being false positives. We use the new pipeline RAVEN to perform a uniform search and validation of transiting planet candidates in TESS data. We focus on a magnitude-limited sample of over 2.2 million main sequence stars well characterised by Gaia and observed by TESS in its Full Frame Images during its first 4 years of operations (sectors 1 to 55). We aim to detect candidates with periods within $0.5-16$ days. RAVEN detects candidates with a box least squares algorithm, classifies them into transiting planets and false positives using machine learning models trained with realistic simulations, and performs statistical validation. We present several samples of candidates with different levels of vetting and validation. We newly validate 118 planets, including 31 newly detected here. We also present a sample of over 2000 candidates not validated but with high probability of being planets, including $\sim1000$ new candidates, a small sample of newly identified mono- and duo-transiting candidates, and a sample of large radii ($>8~\mathrm{R_{\oplus}}$) candidates with high planet probability suited for further follow-up. Our samples of vetted and validated transiting planet candidates represent a major effort towards improving the candidate sample from TESS.

• The paper applies the RAVEN pipeline to TESS-SPOC Full Frame Images from 2.26 million stars, combining GPU-accelerated transit searches, machine learning, astrophysical priors, and model refinement.
• The study validates 143 planets, including 118 newly validated worlds and 31 first-time detections, while producing a vetted catalog of 2,170 candidates, more than 1,000 of them new.
• The results expose important limits of automated validation, including reduced recovery in multi-planet systems and a drop from 220 to 143 validated objects after refined fits, highlighting sensitivity to ephemeris precision and systematic effects.

Overview and motivation

This paper presents a uniform search, vetting, and statistical validation of transiting planet candidates using the RAVEN pipeline (Ranking and Validation of ExoplaNets) applied to TESS Full Frame Image (FFI) light curves from sectors 1–55, corresponding to the first four years of the mission. The stellar sample comprises 2,259,830 main sequence stars brighter than $G = 14$ with Gaia DR3 parallax uncertainties better than 20%, drawn from the Gaia–TESS cross-match catalogue of Doyle et al. The search is restricted to orbital periods of 0.5–16 d and planetary radii up to 16 $R_\oplus$, a parameter space chosen to target the Neptunian desert and its surroundings. The headline results are 143 statistically validated planets — 118 newly validated, including 31 detected for the first time in this work — and a vetted sample of 2,170 candidates, over 1,000 of which are new.

The motivation is twofold. First, the TESS candidate list remains incomplete and heterogeneous; alternative end-to-end searches can recover candidates missed by the official SPOC/QLP TOI pipelines while quantifying the biases at each stage. Second, statistical validation offers an efficient complement to expensive follow-up confirmation, which is impractical for the thousands of candidates produced by all-sky surveys.

Pipeline methodology

RAVEN proceeds through four stages: candidate identification, machine learning classification, probability combination with priors, and parameter refinement.

Candidate identification: Light curves are detrended with a Savitzky-Golay filter (third-degree polynomial, 4-day window), then searched with a GPU-accelerated BLS implementation (cuvarbase) over periods 0.5–16 d. The five highest peaks per star are retained after harmonic de-duplication (peaks at double or half the period of an already-selected signal are skipped), and signals with SDE ≤ 7 or MES ≤ 0.8 are discarded based on injection-recovery tests. This leaves 5,664,552 candidates around 1,812,914 stars.

Classification: RAVEN trains eight binary classifier pairs (one XGBoost GBDT and one Gaussian Process classifier per pair, 16 models total), each distinguishing simulated planets from one false positive scenario: EB, BEB, HEB, HTP, NTP, NEB, NHEB, and NSFP. Crucially, the training sets are built by injecting PASTIS-simulated transits into real TESS-SPOC FFI light curves of stars drawn from the same parent sample, ensuring realistic noise properties. On held-out test sets, the classifiers achieve ROC-AUC scores of 99% (97% for NTP) and precision above 99% at a probability threshold of 0.9. On an independent test set of 1,367 pre-classified TOIs — a deliberately challenging benchmark since these have already passed manual vetting — the pipeline achieves ROC-AUC above 97%, precision above 97% at threshold 0.9, and accuracy of 91%.

Probability combination: Because training sets are balanced, classifier posteriors are combined with priors encoding empirical occurrence rates, detection and recovery probabilities, and a positional probability that the transit occurs on-target rather than on a nearby source (computed via centroid-offset modelling). The final RAVEN probability is the minimum posterior across all planet–FP scenarios, so highly ranked candidates must be probable planets under every scenario considered. A prior cannot be constructed for the NSFP class, which absorbs instrumental noise, stellar variability, and some astrophysical false positives; only the raw classifier posterior is used there.

Parameter refinement: All high-ranked candidates are refit with juliet/dynesty using batman transit models oversampled to account for 30-min and 10-min exposure smearing, a celerite Matérn-3/2 GP for correlated noise, Kipping-parametrised quadratic limb darkening, circular orbits, and wide priors. Stellar radii are taken from Gaia DR2 with a 6% uncertainty assumption.

Vetting and validation criteria

An initial cut on the Planet-NSFP mean classifier probability ($\geq 0.9$) reduces the BLS output to 14,815 candidates, efficiently removing instrumental periodicities near half the sector length (~13.7 d) and low-S/N peaks. Candidates passing both NSFP and RAVEN probability thresholds of 0.9 (3,899 objects) are subjected to further checks:

• removal of 530 candidates with positional probability < 0.5;
• removal of 624 candidates with inter-sector depth variations > 50% or depths < 200 ppm in any sector;
• visual vetting removing 79 period aliases/harmonics;
• removal of 34 candidates with fewer than two full transits;
• removal of 461 candidates with transit S/N < 3.

The resulting vetted sample contains 2,170 candidates around 2,112 stars, over 1,000 of them new. Validation imposes stricter requirements: all four probabilities (initial and juliet-refined, NSFP and RAVEN) ≥ 0.99, at least three full transits, $R_p \leq 8\,R_\oplus$ (to avoid the planet/brown dwarf/low-mass star degeneracy), and a manual check excluding low-S/N or correlated-noise cases. The validated sample contains 143 planets around 140 stars, of which 87 are previously known TOI/CTOI planet candidates (76 PCs, 3 APCs, 8 CTOIs) and 31 are new detections.

A notable methodological finding concerns ephemeris sensitivity: requiring the juliet-refined probabilities removes 77 candidates that passed the initial 0.99 thresholds, reducing the would-be validated sample from 220 to 143. Small changes in fitted ephemerides therefore shift validation probabilities by order-percent amounts, and the authors argue that characterising the intrinsic precision and systematic biases of statistical validation methods should be a priority for the field.

Recovery of known TOIs and CTOIs

Of the 3,098 TOIs within the search range, the BLS recovers 2,633 (84.99%), including 357 KPs, 311 CPs, 1,342 PCs, 175 APCs, 7 FAs, and 441 FPs. Of 1,146 CTOIs, 615 (53.66%) are recovered. Non-recoveries are attributed to shallow transits (the pipeline is reliable only above ~200–300 ppm), low S/N, and competition with deeper signals in multi-planet systems. Multi-candidate recovery is a clear weakness: only about half of multi-candidate TOI systems have more than one candidate recovered, partly because the BLS does not iteratively mask identified signals and because multi-planet architectures are absent from the training simulations. On confirmed-planet TOIs, RAVEN achieves a median final probability of 0.96 overall but only 0.84 for planets in multi-planet systems.

Scientifically notable validated planets

Within the Neptunian desert (Castro-González et al. limits), the work validates five planets: TOI-7008 b ($P \simeq 0.84$ d, $R_p \simeq 4.95\,R_\oplus$), one of the shortest-period desert planets known alongside Kepler-1520 b, LTT 9779 b, TOI-849 b, and TOI-332 b; TOI-5486 b (an M-dwarf host); TOI-4030 b; TOI-2200 b at the ridge boundary; and the CTOI TIC 249022743 b. Because occurrence-rate priors penalise desert planets in the classifiers, validated desert candidates must overcome an inherent bias against them, strengthening confidence in their planetary nature.

In the USP super-Earth regime, the paper validates TIC 18942729 b ($P \simeq 1.00$ d, $R_p \simeq 2.00\,R_\oplus$, found in the third BLS peak), plus TOI-5736 b, TOI-2345 b, TOI-6281 b, and the M-dwarf-hosted CTOI TIC 231949697 b.

Three two-planet systems are validated: TOI-1839 (two sub-Neptunes at ~1.42 d and ~4.02 d), TOI-4156 (~4.46 d and ~12.82 d), and the entirely new system TIC 24750448 (~3.59 d and ~10.14 d). Two additional new vetted candidates are reported in TOI/CTOI host stars (TOI-6484.02 and TIC 142589416.02), the former possibly inducing TTVs on the known outer candidate. Supplementary tables provide four mono-transit candidates and eight visually identified multi-candidate systems with second-planet juliet fits.

Limitations

The authors identify several limitations that bound the interpretation of their samples:

• Harmonics: the fixed 0.5–16 d search window means signals outside this range can appear as aliases; despite visual vetting, complex harmonics persist in the vetted sample, particularly when data gaps admit spurious shorter periods. Conversely, if a harmonic is selected first by the BLS, the true peak may be discarded.
• Multi-candidate systems: no multi-planet training simulations, no TTV handling, and no iterative masking depress recovery rates, as demonstrated by the TOI statistics above.
• Depth sensitivity: SDE/MES cuts, the 200 ppm depth floor, and S/N ≥ 3 requirement limit sensitivity to small planets; the bulk of vetted candidates exceed ~1.2 $R_\oplus$.
• Crowded fields: the positional-probability cut (> 0.5) likely reduces sensitivity in crowded regions, though the dependence is entangled with transit S/N.
• Uniform fits: single-planet, circular-orbit juliet fits with identical priors trade accuracy for homogeneity; comparison against NEA values shows Pearson correlations of 1.0 for periods (ratio scatter ~0.0001) and 0.97 for radii (ratio scatter < 0.1), but tailored fits could improve individual systems.
• Validation status: the authors explicitly caution that statistically validated planets should not be treated as equivalent to mass-confirmed planets; many of the new validations lack any follow-up observations.

Conclusion

Applying RAVEN uniformly to ~4 years of TESS-SPOC FFI photometry of 2.26 million Gaia-characterised main sequence stars yields a tiered set of products: 14,815 initially vetted TCEs, 2,170 fully vetted candidates (~1,000 new), 143 statistically validated planets (118 newly validated, 31 new detections), 207 high-probability large-radius candidates reserved for follow-up, four mono-transit events, and eight visually flagged multi-candidate systems. Because each pipeline step has been characterised via realistic injections, the vetted sample is suitable for occurrence-rate inference, which the authors defer to companion work. The principal open questions left by this study are the quantitative precision and systematics of statistical validation itself — highlighted by the 77-candidate discrepancy between BLS-based and juliet-refined probabilities — and improved treatment of multi-planet systems within automated vetting frameworks.