🔭 astrophysics

ExoNet: Multimodal Deep Learning for TESS Exoplanet Candidate Identification via Phase-Folded Light Curves, Stellar Parameters, and Multi-Head Attention Fusion

This paper introduces ExoNet, a multimodal deep learning framework that combines phase-folded light curves and stellar parameters via 1D CNNs and multi-head attention to effectively automate the identification and validation of exoplanet candidates in TESS data.

Original authors: Md. Rashadul Islam

Published 2026-04-20

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Md. Rashadul Islam

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 night sky as a massive, bustling city, and astronomers are like detectives trying to find tiny, invisible ghosts (planets) that occasionally walk past the streetlights (stars). When a ghost walks in front of a light, the light dims just a tiny bit. NASA's TESS satellite is a super-powered security camera that has been watching this city for years, capturing millions of these tiny "dimming" moments.

But here's the problem: The camera is too good. It has found over 7,800 potential ghosts, but human detectives have only confirmed 720 of them. The rest are stuck in a "waiting room" because checking them one by one by hand is too slow. It's like having a library with millions of books but only one librarian who can read one page a day.

This paper introduces ExoNet, a new "AI detective" designed to speed up the process. Here is how it works, explained simply:

1. The Three-Pronged Detective (Multimodal Learning)

Old AI detectives usually looked at just one thing: the graph of the light dimming (the "light curve"). It's like trying to identify a suspect only by their shadow.

ExoNet is smarter. It acts like a detective who gathers three different types of clues at once:

The "Wide-Angle" View: It looks at the entire orbit of the star to see the big picture.
The "Zoom-In" View: It zooms in tight on the exact moment the light dims to see the shape of the "ghost's" walk.
The "ID Card" View: It checks the star's personal file (temperature, size, age).

The Analogy: Imagine you are trying to find a specific person in a crowd.

The Old AI only looks at the person's silhouette.
ExoNet looks at the silhouette, zooms in on their face, and checks their ID card to see if their height and age match the description. By combining all three, it's much harder to fool.

2. The "Attention" Mechanism

The paper mentions "Multi-Head Attention." In plain English, this is like giving the AI detective multiple pairs of glasses.

One pair of glasses focuses on the start of the dimming.
Another pair focuses on the end.
Another pair looks for weird glitches in the data.

Instead of getting overwhelmed by all the data, the AI learns to ignore the "noise" (like a bird flying in front of the camera) and focus only on the parts that look like a real planet. It's like a seasoned chef who can instantly taste a dish and ignore the salt shaker, focusing only on the flavor of the soup.

3. The Results: Finding the "Gold"

The researchers trained this AI using data from the Kepler mission (the previous generation of planet hunters) and then let it loose on the 200 most recent, unconfirmed TESS candidates.

The Scorecard:

Accuracy: The AI got about 81% of the test cases right, which is a huge improvement over looking at just one type of data.
The Big Catch: Out of 200 candidates, it flagged 35 as "High Confidence."
The "Holy Grail": It found 19 candidates that are in the "Goldilocks Zone" (not too hot, not too cold—just right for liquid water).

4. The Star of the Show: TOI-7949.87

The most exciting discovery is a planet candidate named TOI-7949.87.

Size: It is almost exactly the size of Earth (0.97 times our size).
Host: It orbits a star very similar to our Sun.
Confidence: The AI is 99.64% sure this is a real planet.

The Metaphor: If the other candidates are like "maybe it's a ghost, maybe it's a trick of the light," TOI-7949.87 is the AI pointing at a suspect and saying, "I'm 99% sure that's the one. Go check their ID right now."

Why Does This Matter?

Before ExoNet, humans had to stare at graphs for hours to decide if a signal was real. Now, this AI can do in two hours what used to take weeks of human effort.

It doesn't replace the human scientists; it acts as a super-assistant. It filters out the noise and hands the human team a short list of the most promising candidates, specifically pointing out the ones that might be habitable.

In summary: ExoNet is a smart, multi-sensory AI that combines different types of data to find Earth-like planets faster than ever before, giving us a better chance to answer the question: "Are we alone in the universe?"

1. Problem Statement

NASA's Transiting Exoplanet Survey Satellite (TESS) has identified over 7,800 planet candidates as of early 2026, yet fewer than 720 have been confirmed. This creates a massive "validation bottleneck." Traditional vetting relies on manual inspection of Data Validation (DV) reports, a process that cannot scale to the volume of data. While previous machine learning solutions (e.g., AstroNet, ExoMiner) have utilized Convolutional Neural Networks (CNNs), they often treat light curves and stellar parameters separately or lack advanced attention mechanisms to jointly learn from both modalities. There is a need for a scalable, automated, multimodal framework that can efficiently vet candidates and prioritize those in the habitable zone.

2. Methodology: The ExoNet Architecture

ExoNet is a trimodal deep learning framework designed to jointly encode phase-folded light curves and stellar parameters. The architecture consists of three parallel streams fused via a late-fusion mechanism:

Data Inputs:
- Global Light Curve View: A phase-folded light curve covering the full orbit ( $N=2001$ bins, $\phi \in [-0.5, 0.5]$ ) to capture secondary eclipses and asymmetries.
- Local Light Curve View: A zoomed-in view of the transit event ( $N=2001$ bins, $|\phi| < 2\Delta t/P$ ) to preserve ingress-egress shapes.
- Stellar Parameters: A vector of 8 scalar features per target (Period, Duration, Depth, Radius, Equilibrium Temp, Effective Temp, Surface Gravity, Metallicity).
Network Components:
- 1D CNN Encoders: Two separate 1D CNN branches (Global and Local) process the light curves. Each uses five residual Conv1D blocks with increasing filter counts (16 to 256) and stride-2 pooling, reducing the sequence length from 2001 to ~62 steps.
- Multi-Head Attention (MHA): An 8-head Multi-Head Attention mechanism is applied over the temporal feature maps of the CNN outputs. This allows the model to dynamically upweight critical regions (e.g., transit ingress/egress) and downweight systematics, mimicking human vetting behavior.
- Stellar MLP: A Multi-Layer Perceptron processes the 8-dimensional stellar vector, outputting a 128-dimensional embedding.
- Late Fusion Head: The embeddings from the two CNN branches (256-d each) and the Stellar MLP (128-d) are concatenated into a 640-dimensional vector. This is passed through a residual fusion MLP (512 $\to$ 256 units) with a linear shortcut connection before a binary classification head (Softmax).
Training Protocol:
- Dataset: Trained on labeled Kepler data (NASA Exoplanet Archive) with a 70/15/15 train/val/test split.
- Augmentation: Gaussian noise, phase shifts, and flux scaling.
- Optimization: AdamW optimizer with OneCycleLR scheduling, trained on a single NVIDIA L4 GPU using mixed-precision (FP16) in under 2 hours.

3. Key Contributions

Trimodal Architecture: First application of a 1D CNN + Multi-Head Attention + Residual Late-Fusion architecture for exoplanet vetting, explicitly fusing global views, local views, and stellar context.
Ablation Studies: Demonstrated that each modality (Global, Local, Stellar) and the MHA mechanism provides measurable performance gains.
TESS Candidate Catalog: Application of the model to 200 unconfirmed TESS Planet Candidates (PCs), identifying high-confidence signals.
Discovery of TOI-7949.87: Identification of a near-Earth radius candidate ( $0.97 R_\oplus$ ) around a Sun-like host with 99.64% confidence.
Open Science: Full release of the code, trained models, and a ranked catalog of candidates, including 19 habitable-zone signals.

4. Results

Performance on Kepler Test Set:
- Test AUC: 0.8112 (Validation AUC: 0.8566).
- Accuracy: 72.86%.
- Ablation Impact:
  - Global CNN only: 0.749 AUC.
  - Global + Local CNN: 0.793 AUC.
  - Full Trimodal (ExoNet): 0.8112 AUC.
  - The addition of Multi-Head Attention contributed approximately +0.013 AUC over global average pooling alone.
TESS Candidate Analysis (200 PCs):
- Identified 35 high-confidence candidates ( $\ge 70\%$ probability).
- 27 candidates exceeded 85% confidence.
- 19 candidates fall within the habitable zone temperature range (200–400 K).
- Top Candidate: TOI-7949.87 (99.64% probability, $0.97 R_\oplus$ , 15-day period, Sun-like host).
- Notable Follow-up Target: TOI-5521.20 (M-dwarf host, $1.90 R_\oplus$ , 35-day period, $327$ K), suitable for radial velocity follow-up.

5. Significance and Limitations

Significance:
- Scalability: The model trains in under two hours on a single GPU, offering a highly accessible alternative to proprietary pipelines like ExoMiner.
- Interpretability: The use of Multi-Head Attention allows the model to learn where to look in the light curve, replicating human vetting logic without explicit rule coding.
- Habitable Zone Focus: The pipeline successfully prioritizes candidates in the habitable zone, aiding future JWST atmospheric characterization efforts.
Limitations:
- Data Gaps: ~1/3 of the TESS candidates lacked available light curves, requiring zero-padding.
- Domain Shift: The model was trained on Kepler data without explicit domain adaptation for TESS, though transfer learning was implied in the architecture's robustness.
- Habitability Proxy: Equilibrium temperature is used as a rough proxy; atmospheric modeling is required to confirm true habitability.

Conclusion

ExoNet represents a significant step forward in automated exoplanet vetting by effectively combining multimodal data fusion with attention mechanisms. It successfully bridges the gap between raw TESS data and high-confidence planet candidates, providing a reproducible, open-source tool for the astronomical community to prioritize targets for confirmation and characterization.