T16 Project: Exoplanet Search & ML Techniques

• T16 Project is a large-scale initiative that applies advanced image-subtraction photometry to TESS full-frame images, enabling exoplanet transit detection in faint stars.
• It integrates rigorous astrophysical data processing with machine learning, achieving high precision and robust classification metrics such as an F1 score near 0.91.
• The project produces an extensive catalog of over 83 million light curves, significantly expanding the exoplanet candidate yield and offering insights into both transit science and speech quality assessment.

The T16 Project encompasses a series of large-scale, high-precision astrophysics and machine learning initiatives, with its primary instantiation being an exoplanet transit search leveraging image-subtraction photometry from NASA TESS full-frame images, as well as its application in speech quality assessment in the AudioMOS Challenge. The T16 Project's astrophysical phase delivers the most comprehensive set of full-frame time-series light curves (over 83 million) for stars with TESS magnitudes down to T=16, pushing planetary census efforts into sparsely explored faint regimes. In its machine learning context, T16 denotes the first MambaRate “few-shot” submission for cross-sampling-rate speech MOS assessment. This article focuses on the scientific objectives, methodologies, outcomes, and implications of the T16 Project in exoplanetary science, also noting connections to broader computational methodologies.

1. Scientific Motivation and Objectives

The T16 Project addresses the longstanding limitation of TESS data products, whose main pipelines (SPOC, QLP) have historically prioritized relatively bright stars (T ≲ 13.5). Statistical yield predictions for transiting exoplanets, however, indicate a significant, untapped population around much fainter hosts. To realize the scientific return of TESS Cycle 1 FFIs—which surveyed ∼100 million stars to T<16 across the entire sky—T16 constructs uniformly detrended difference-image light curves for all accessible targets. The project aims to expand the exploration volume for exoplanet detection, mitigate crowding and background-related systematics that hampered aperture photometry, and enable statistical demographic studies of planets around metal-poor, low-luminosity, and otherwise under-investigated hosts (Roth et al., 20 Apr 2026, Hartman et al., 19 Feb 2025).

2. Data Acquisition, Image Subtraction, and Light-Curve Detrending

T16 processed Cycle 1 FFIs (sectors 1–13), acquired with 30-minute cadence and 16 detectors per sector, through the CDIPS (Cluster Difference Imaging Photometric Survey) image-subtraction pipeline:

• Construction of high-S/N reference frames from ~50 aligned images, followed by convolution with sector-specific kernels to optimally subtract science frames, as per Alard & Lupton (1998).
• Photometry is extracted via circular apertures, with PSF diagnostics encoded, and Gaia DR2 is used for flux calibration to mitigate crowding effects.
• Time stamping utilizes barycentric corrections recomputed with VARTOOLS+JPL-Horizons, yielding ≤0.04 s RMS timing accuracy.

Systematics are suppressed via:

1. Spline-based External Parameter Decorrelation (SEPD), fitting basis splines over time and linear decorrelations against detector positions and CCD temperature.
2. Trend Filtering Algorithm (TFA), applying a basis of 200 non-variable template stars to remove residual correlated noise.

The resulting SEPD+TFA light curves achieve scatter levels at or near the theoretical noise floor for $T\sim8-15$ mag, outperforming public alternatives for faint stars (see Section 4) (Hartman et al., 19 Feb 2025).

3. Automated Transit Detection and Machine Learning Classification

With over $10^8$ systematics-corrected light curves, T16 employs a semi-automated transit search and classification pipeline:

• CETRA Implementation: GPU-accelerated matched-filter searches identify single-transit candidates and fold over trial period grids. The key detection statistic, $S/\mathrm{PN}$, quantifies signal significance as a function of white and red noise components.
• Feature Extraction: Fixed-period BLS (box least squares) analyses probe period harmonics, extracting 53 descriptive features per candidate, including transit depth, duration, S/PN at harmonics, $\Delta\chi^2$ improvements, astrophysical parameters from Gaia, etc.
• Random Forest Classification: Dual random-forest classifiers, stratified by stellar brightness, assign candidates to "Planet Transits," "Eclipsing Binaries," or "Non-Detections." These models achieve $\mathrm{F}_1\approx0.91$ and ROC AUC ≈ 0.98–0.99 on hold-out data.
• Vetting Criteria: Candidates require $p_{\text{transit}} \geq 0.5$, $S/\mathrm{PN}>10$, spatial blend rejection, BATMAN-fit parameter consistency ($R_p/R_\star\leq0.18$), and clean subtraction images. Final inspection includes manual vetting of ∼50,000 high-S/N candidates (Roth et al., 20 Apr 2026).

4. Catalog Release, Photometric Performance, and External Comparisons

The T16 data product consists of over 83 million light curves for 54 million+ stars (Cycle 1), with organization by sector, detector, and full metadata epochs. Quantitative comparisons to QLP, SPOC, TGLC, and GSFC pipelines demonstrate:

Compared pipeline	Fraction of T16 light curves with lower MAD* (detrended)
QLP KSPSAP	78.0%
SPOC PDCSAP	69.8%
TGLC calibrated aperture	85.6%
GSFC corrected	43.1%

*(Median Absolute Deviation, per-pipeline matched comparison in relevant magnitude ranges.)

For $T>13.5$, T16 offers substantially superior scatters, extending precision transit search capability to ∼16th mag, a regime previously omitted by public reductions. SEPD+TFA light curves reliably recover 89% of known TOI periods; undetrended (IRM) curves support studies of long-period and high-amplitude variables, such as Cepheids and OGLE eclipsing binaries (Hartman et al., 19 Feb 2025).

5. Exoplanet Candidate Yield and Vetting

From the initial candidate pool subjected to automated and manual vetting, the T16 project catalogues:

• 11,554 planet candidates (11,143 unique multi-transit systems; 10,091 are newly identified by T16; 1,052 are previously known TESS candidates).
• 411 single-transit events (without period constraints).
• Parameterizations for multi-transit candidates include ephemerides and MCMC-derived uncertainties.

The population is dominated by large-radius objects ($>0.5\,R_J$), with ∼170 Neptunians ($10^8$0) and 11 sub-2 $10^8$1 super-Earths. Approximately 70% orbit stars with $10^8$2, and ∼6% are associated with hosts of $10^8$3 (Roth et al., 20 Apr 2026).

6. Radial Velocity Confirmation and Scientific Significance

To validate the pipeline, T16 secured Magellan/PFS radial velocities for TIC 183374187, confirming a transiting hot Jupiter ($10^8$4 d, $10^8$5, $10^8$6) orbiting a metal-poor, $10^8$7-enhanced thick-disk star. Blend and false positive scenarios are decisively excluded.

By expanding the exoplanet census from ∼7,800 to ∼19,400 TESS planet candidates, T16 reveals abundance of large planets at faint magnitudes and opens parameter space for metal-poor hosts, ultra-short-period planets, and rare evolutionary scenarios. These results realize predictions from the literature that thousands of faint transiting planets awaited discovery with advanced pipelines (Roth et al., 20 Apr 2026).

7. Public Data Access, Limitations, and Future Directions

The T16 light curves, including IRM, EPD, and TFA products, are publicly available as High-Level Science Products (HLSP) via MAST. Noted limitations include suppression of >1 day astrophysical signals in detrended products and residuals in certain sector/CCD combinations. Flux calibration from Gaia DR2 has intrinsic ∼0.1 mag uncertainty.

Planned future work includes processing additional TESS cycles, extending search sensitivity to longer orbital periods, systematic injection–recovery validation, further false positive analysis (e.g., TRICERATOPS statistical validation tools), and deep learning classification to reduce the vetting bottleneck. The experience in automated transit recovery and machine learning enabled by T16 also informs data-mining strategies more broadly in time domain astronomy (Hartman et al., 19 Feb 2025, Roth et al., 20 Apr 2026).