🔭 astrophysics

ExoMiner++ 2.0: Vetting TESS Full-Frame Image Transit Signals

This paper demonstrates that ExoMiner++ 2.0, an adapted machine learning framework, successfully generalizes to TESS Full-Frame Image data to robustly distinguish planetary transits from false positives and artifacts, thereby enabling large-scale vetting of the full TESS dataset for future population studies.

Original authors: Miguel J. S. Martinho, Hamed Valizadegan, Jon M. Jenkins, Douglas A. Caldwell, Joseph D. Twicken, Ben Tofflemire, Marziye Jafariyazani

Published 2026-02-02

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Miguel J. S. Martinho, Hamed Valizadegan, Jon M. Jenkins, Douglas A. Caldwell, Joseph D. Twicken, Ben Tofflemire, Marziye Jafariyazani

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the TESS space telescope as a giant, high-powered security camera scanning the entire sky. Its job is to watch millions of stars, looking for tiny, rhythmic dips in their brightness. These dips are like a shadow passing over a lightbulb, which usually means a planet is crossing in front of its star.

However, TESS has two different "modes" of watching:

The "2-Minute Mode": It focuses on a small, pre-selected list of about 20,000 favorite stars, taking a snapshot every 2 minutes. This is like a security guard watching a VIP list very closely.
The "Full-Frame Image (FFI) Mode": It takes a picture of the entire sky every 200 seconds (or 10-30 minutes, depending on the time). This is like a wide-angle security camera watching the whole neighborhood. It sees way more stars, but the pictures are grainier, and the "guards" (the data processing software) have a harder time telling if a shadow is a planet or just a trick of the light.

The Problem: Too Many False Alarms

Because the "Full-Frame" mode sees so many stars, it finds thousands of potential planet signals. But many of these are false alarms.

Some are caused by background stars that happen to line up with the target star.
Some are caused by binary stars (two stars orbiting each other) that eclipse one another.
Some are just noise or glitches in the camera.

In the past, scientists had to manually check these signals, which is like a human security guard trying to review millions of hours of footage. It's slow and exhausting.

The Solution: ExoMiner++ 2.0 (The "Super-Inspector")

The authors of this paper built a new version of an AI detective called ExoMiner++ 2.0. Think of this AI as a super-smart intern who has been trained to look at the "footage" (the light curves and images) and instantly decide: "Is this a real planet, or is it a fake?"

Here is how they made this new AI smarter than the old one:

1. Teaching it to "See" the Neighborhood
The old AI looked at the target star in isolation. But in the crowded "Full-Frame" images, other stars often sit right next to the target. If a neighbor star has a planet, its shadow might look like it belongs to the target star.

The Upgrade: The new AI now gets a special "neighborhood map" as an extra input. It sees a picture of the target star plus all the nearby stars, their brightness, and their exact locations. It's like giving the security guard a map of the whole block, not just the house they are watching. This helps the AI realize, "Wait, that shadow is actually coming from the house next door, not the one we are watching."

2. Sharper Eyes
The AI now looks at the "difference images" (photos showing what changed between snapshots) with higher resolution. It's like switching from a blurry 480p camera to a crisp 4K camera, allowing it to spot tiny shifts in where the shadow is coming from.

3. Training on Everything
The AI was trained using data from both the "VIP list" (2-minute data) and the "whole neighborhood" (FFI data). By learning from the clean, high-quality VIP data, it became better at spotting patterns in the messy, noisy neighborhood data. It's like a student who studies perfect textbook examples first, then applies that knowledge to solve messy real-world problems.

What Did They Find?

The team tested this new AI on over 150,000 unverified signals from the "Full-Frame" mode.

It works well: The AI successfully distinguished real planets from fake signals with about 95% accuracy in a specific statistical test.
It's cautious: The new AI is slightly more "conservative." If a signal looks a bit suspicious (like it might be coming from a neighbor star), the AI gives it a lower score. This means it might miss a few real planets to avoid calling a fake one a planet. The authors prefer this because it saves scientists time by filtering out the obvious fakes.
The Result: They produced a new "shortlist" of the most promising candidates. These are the signals that passed the AI's rigorous test and are now ready for human scientists to double-check and confirm.

The Bottom Line

This paper is about building a better filter for the TESS telescope's massive data dump. By teaching the AI to understand the context of the stars around its target, they created a tool that can handle the messy, crowded data of the "Full-Frame" mode. This allows the exoplanet community to focus their time and resources on the most likely new planets, rather than getting bogged down by thousands of false alarms.

Technical Summary: ExoMiner++ 2.0: Vetting TESS Full-Frame Image Transit Signals

Problem Statement
The Transiting Exoplanet Survey Satellite (TESS) Full-Frame Images (FFIs) provide photometric time series for millions of stars, significantly expanding the search space beyond the pre-selected 2-minute targets. However, FFI data present unique challenges for transit identification and vetting, including lower cadence (ranging from 200s to 30 minutes depending on the mission phase), higher levels of flux contamination from neighboring stars, and a lack of extensive multi-sector coverage compared to 2-minute data. The resulting Threshold Crossing Events (TCEs) require automated, scalable vetting systems capable of distinguishing planetary signals from astrophysical false positives (e.g., eclipsing binaries, background objects) and instrumental artifacts with minimal false positive rates. Existing vetting frameworks, such as the original ExoMiner++, were primarily trained on 2-minute data and lacked specific mechanisms to handle the contamination and spatial resolution issues inherent to FFI data.

Methodology
The authors present ExoMiner++ 2.0, an adaptation of the ExoMiner++ deep learning framework designed to vet TCEs from SPOC FFI light curves. The methodology involves three primary components:

Data Curation and Labeling:
- The dataset comprises TCEs from SPOC 2-minute data (Sectors 1–88) and FFI data (Sectors 36–72).
- Labels are derived from the ExoFOP TOI catalog, Prsa's TESS eclipsing binary (EB) catalog, and TEC flux triage results. TCEs are classified into binary labels: "planets" (Confirmed/Kepler Planets) and "non-planets" (Brown Dwarfs, EBs, False Positives, Non-Transiting Phenomena, and Unknowns).
- The dataset includes 268,787 2-minute TCEs and 222,054 FFI TCEs.
Architectural and Preprocessing Modifications:
- Neighboring Star Encoding: A novel image-based representation is introduced to encode the relative location and brightness of known neighboring TIC stars. This information is mapped to the CCD pixel frame and added as an extra input channel to the 'Difference Image' branch. This allows the model to identify off-target transit sources and contamination-driven false positives, particularly in crowded fields.
- Resolution and Branch Adjustments: The difference image resolution is increased (sampling factor of 5, resulting in 55×55 px) to better resolve photocenter offsets. The 'Unfolded Flux' branch is merged with other transit-view flux branches. The 'Momentum Dump' branch is removed as it was found non-informative.
- Model Architecture: The model incorporates global average pooling, learned downsampling (replacing max pooling), batch normalization, skip connections (residual blocks), and layer normalization to improve training stability and reduce overfitting.
Training Strategy:
- The model employs multi-source learning, training simultaneously on both 2-minute and FFI datasets. This approach allows the model to learn shared representations across different cadences and noise characteristics, leveraging the richer labels and cleaner systematics of the 2-minute data to improve generalization to FFI data.
- Experiments utilize 5-fold cross-validation with target-based splitting to prevent data leakage.

Key Results

Performance: ExoMiner++ 2.0 achieves a Precision-Recall Area Under the Curve (PR AUC) of approximately 0.95 under multi-source training. While the model shows slightly lower global recall compared to the original ExoMiner++ (0.865 vs. 0.912 at threshold 0.5), it demonstrates improved performance in identifying specific negative subclasses, particularly False Positives (FP) and Brown Dwarfs (BD).
Contamination Handling: The inclusion of neighboring star data makes the model more conservative. It systematically lowers scores for transit-like signals originating near bright neighbors, effectively reducing confidence in potential false positives caused by off-target sources.
Data Source Comparison: Performance on 2-minute data remains higher than on FFI data (PR AUC 0.967 vs. 0.952). The authors attribute the lower FFI performance to higher flux contamination (lower CROWDSAP metrics), greater variability, and fewer multi-sector runs, which limit the detection of long-period signals.
Vetting Catalog: The model was applied to 151,616 unlabeled FFI TCEs. The resulting catalog identifies 2,831 candidates with a score > 0.5, including 447 high-confidence candidates (score > 0.95). The distribution of high-scoring candidates shows a cluster consistent with hot Jupiters (periods 2–5 days, radii 10–20 $R_\oplus$ ).

Significance and Claims
The paper claims that ExoMiner++ 2.0 successfully extends the applicability of deep learning vetting to the full TESS FFI dataset. Its primary contributions are:

First Integration of Neighbor Data: It is the first TESS machine-learning vetter to explicitly encode the relative magnitude and location of neighboring stars as an image representation, improving the identification of off-target transit sources.
Robust Generalization: The model demonstrates effective generalization beyond the canonical 2-minute domain, providing robust discrimination between planetary signals, astrophysical false positives, and instrumental artifacts in the more challenging FFI domain.
Scalable Vetting: By automating the vetting of hundreds of thousands of FFI TCEs, the work provides a prioritized, ranked catalog of candidates. This supports future population studies and enables the exoplanet community to focus follow-up resources on a smaller, cleaner set of candidates, including those orbiting stars not previously targeted by 2-minute observations.

The authors note limitations, including the 21" pixel scale of TESS which limits the separation of on-target and nearby sources even with enhanced resolution, and the relatively small size of the labeled FFI dataset for rare subclasses. They suggest that further improvements could be achieved through dedicated hyperparameter optimization of the 'Difference Image' branch and the incorporation of attention mechanisms.

The Problem: Too Many False Alarms

The Solution: ExoMiner++ 2.0 (The "Super-Inspector")

What Did They Find?

The Bottom Line

More like this