🔭 astrophysics

ASTRAFier: A Novel and Scalable Transformer-based Stellar Variability Classifier

This paper introduces ASTRAFier, a novel, scalable Transformer-based model that combines BiLSTM and CNNs to classify stellar variability directly from raw light curves without feature engineering, achieving high accuracy on Kepler and TESS data and successfully processing nearly 2.8 million TESS light curves to release a new variability catalog.

Original authors: Paul F. X. Gregory, Jeroen Audenaert, Mykyta Kliapets, Daniel Muthukrishna, Andrew Tkachenko, Marek Skarka, George R. Ricker

Published 2026-04-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Paul F. X. Gregory, Jeroen Audenaert, Mykyta Kliapets, Daniel Muthukrishna, Andrew Tkachenko, Marek Skarka, George R. Ricker

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 where every star is a person with a unique personality. Some stars are calm and steady, others pulse like a heartbeat, some spin wildly, and others are actually two stars dancing around each other, eclipsing one another. For decades, astronomers have been trying to take a census of this city, but the sheer number of stars is overwhelming.

Enter ASTRAFier, a new digital detective introduced in this paper. It's a super-smart computer program designed to look at the "light curves" of stars (graphs showing how bright a star gets and dims over time) and instantly tell us what kind of star it is.

Here is the story of how ASTRAFier works, explained without the heavy math.

1. The Problem: Too Much Data, Too Little Time

Space missions like Kepler and TESS have been taking photos of the sky non-stop. They've collected millions of light curves. If you tried to look at them all by hand, it would take a human lifetime. Even old computer programs needed humans to first "translate" the light curves into a list of numbers (like calculating the average brightness or the rhythm of the pulse) before the computer could guess the star's type. This was slow and often missed the subtle details.

2. The Solution: A "Super-Brain" for Stars

The authors built ASTRAFier, which is a Transformer-based model. To understand what that means, think of it like this:

The Old Way (Feature Engineering): Imagine trying to identify a song by first writing down the tempo, the key, and the volume, and then feeding those numbers into a calculator.
The ASTRAFier Way (End-to-End Learning): ASTRAFier is like a music lover who just listens to the raw audio. It doesn't need you to explain the rules of music; it learns the patterns itself by listening to millions of examples.

3. How ASTRAFier Thinks: The Three Musicians

ASTRAFier isn't just one brain; it's a trio of specialized musicians working together in a band to solve the mystery of a star.

The BiLSTM (The Time Traveler): This part of the model is great at remembering the past and predicting the future. It looks at a star's brightness history and says, "Okay, it was dim here, bright there, and now it's dim again. I see a pattern." It handles the sequence of events.
The CNN (The Pattern Spotter): This is like a detective with a magnifying glass. It scans the light curve for specific shapes and local details—like a sharp spike (an eclipse) or a smooth wave (a pulsation). It ignores the big picture to focus on the tiny, crucial details.
The Transformer (The Context Master): This is the star of the show. It uses "Self-Attention." Imagine you are reading a long story. The Transformer is the reader who can instantly jump back to the first page to remember a character's name while reading the last page. It connects the beginning of a star's light curve to the end, understanding how the whole story fits together, even if the story is very long or has gaps (which happens in space data when satellites take breaks).

The Secret Sauce: These three musicians play together in a loop. They pass their findings back and forth, refining their understanding until they are confident about what they are seeing.

4. The Training: School for Stars

To teach ASTRAFier, the scientists showed it a massive library of "labeled" stars.

The Textbooks: They used data from the Kepler mission (where the data is very clean and high-quality) and TESS (which covers the whole sky but is a bit noisier).
The Lesson: They showed the model millions of light curves and told it, "This one is a pulsating star," "This one is an eclipsing binary," etc.
The Result: ASTRAFier learned to recognize the "fingerprint" of each star type.
- On the clean Kepler data, it got 94% of the answers right.
- On the messier, real-world TESS data, it still got 88% right.

5. The Big Mission: Classifying 2.8 Million Stars

Once the model was trained, the scientists unleashed it on 2.8 million light curves from the TESS mission. It did the job in a fraction of the time it would take a human.

What did they find?

They created a massive catalog of variable stars.
They found that the model is very good at spotting eclipsing binaries (stars that block each other's light).
However, the model sometimes gets confused between stars that spin (rotational variables) and stars that are just irregular (aperiodic), especially if the spin is very slow.
They also found that the model sometimes mistakes "instrument noise" (glitches from the telescope) for real stars, so they added a special "Junk" category to catch the errors.

6. Why This Matters

ASTRAFier is a game-changer because:

It's Fast: It can process millions of stars quickly, allowing astronomers to find the most interesting targets for follow-up studies.
It's Flexible: Because it learns directly from the raw data, if we discover a new type of star behavior in the future, we can just retrain the model without having to invent new mathematical formulas.
It's Scalable: As space missions get bigger and collect even more data (like the upcoming PLATO mission), ASTRAFier can grow with them.

In Summary:
ASTRAFier is like a highly trained, tireless intern who can listen to the "heartbeat" of 2.8 million stars, figure out their personalities, and file them away correctly, all while learning from its mistakes and getting smarter every day. It turns a mountain of data into a map of the universe's living, breathing stars.

1. Problem Statement

Space missions like Kepler and TESS have generated millions of high-quality light curves, creating an unprecedented opportunity to study stellar variability. However, the sheer scale of this data necessitates efficient, automated analysis methods. Traditional approaches rely heavily on feature engineering (extracting statistical moments, periodograms, entropy, etc.) followed by classifiers like Random Forests or Gradient Boosting. These methods are often labor-intensive, require domain expertise to design features, and may fail to capture complex temporal dependencies inherent in stellar variability. Furthermore, existing deep learning solutions often struggle with the variable lengths of astronomical time series and the presence of data gaps.

2. Methodology: The ASTRAFier Architecture

The authors introduce ASTRAFier (Astronomical Sequence TRansformer-based vAriability classifier), a novel end-to-end deep learning framework that processes raw light curves directly without manual feature engineering. The architecture is a hybrid sequential model integrating three core components:

Bidirectional LSTM (BiLSTM): Processes the input sequence in both forward and backward directions to capture sequential memory and long-term dependencies. It handles the temporal context of the light curve.
Transformer Encoder: Utilizes Multi-Head Self-Attention (MHSA) to capture global context and long-range correlations between time steps.
- Innovation: The authors modify standard positional encodings to be time-dependent rather than index-dependent. This accounts for the irregular time sampling and gaps common in astronomical data (e.g., spacecraft operations).
Convolutional Neural Networks (CNNs):
- CNN Projection: A block following the BiLSTM that reduces dimensionality while preserving bidirectional features using 1-D convolutions.
- CNN Module: Applied after the Transformer, using multi-scale kernels (sizes 3, 7, 15, and 111) to detect localized temporal patterns and trends at various receptive fields.
Residual Connections: "Add and Norm" blocks connect the output of each module back to its input, ensuring gradient flow and preserving characteristic information extracted by each layer.
Variable-Length Handling: The model employs binary padding masks to ignore padded positions during attention calculations and averaging, ensuring that artificial padding does not influence the learned representations.

Output Layer: The sequence output is aggregated via masked global average pooling (summing only valid time steps) and passed through a Multi-Layer Perceptron (MLP) with a softmax activation to produce class probabilities. Monte-Carlo dropout is used during inference to estimate predictive uncertainty.

3. Key Contributions

End-to-End Learning: ASTRAFier eliminates the need for manual feature extraction, operating directly on raw flux time series.
Hybrid Architecture: It successfully integrates the complementary strengths of BiLSTMs (sequential memory), Transformers (global attention), and CNNs (local feature detection) into a single scalable model.
Robustness to Data Irregularities: The custom time-dependent positional encoding and masking strategies allow the model to handle the irregular cadence and gaps typical of TESS and Kepler data.
Scalability: The model is designed for massive deployment, successfully classifying millions of light curves with high efficiency.
Public Release: The authors release a comprehensive catalog of variable star candidates and the trained model code.

4. Results

The model was trained and validated using a combination of Kepler and TESS (Quick-look Pipeline, QLP) light curves across 8 variability classes (e.g., Eclipsing Binaries, $\delta$ Scuti, RR Lyrae, Solar-like oscillators, etc.).

Performance on Kepler: Achieved a classification accuracy of 94.26% on the holdout set, comparable to state-of-the-art ensemble methods but with a simpler architecture.
Performance on TESS: Achieved 88.22% accuracy on TESS data when trained on a combined Kepler+TESS dataset.
Class-Specific Performance:
- Eclipsing Binaries: 100% recall on the holdout set.
- RR Lyrae/Cepheids: High precision (>92%), though recall varied slightly due to small sample sizes.
- Solar-like Oscillators: Demonstrated strong performance, though some confusion with g-mode pulsators was noted in the HR diagram analysis.
Ablation Study: Removing any of the three core modules (BiLSTM, Attention, or CNN) resulted in a drop in performance, confirming that each component contributes uniquely to the model's success.
Scalability: The model classified ~2.8 million TESS light curves (Sectors 14, 15, and 26) in approximately 8 hours using two NVIDIA H200 GPUs.

5. Significance and Future Work

Efficiency: ASTRAFier offers a computationally efficient alternative to complex ensemble methods, enabling rapid classification of the entire TESS dataset and future missions like PLATO.
Catalog Generation: The deployment resulted in a new variability catalog for the Kepler Field-of-View in TESS sectors, identifying candidates for follow-up observations.
Foundation Model Potential: The authors highlight that this architecture serves as a stepping stone toward self-supervised foundation models for astronomy, which could leverage unlabeled data for tasks beyond classification, such as anomaly detection and parameter estimation.
Limitations & Future Directions: The authors note that performance is limited by the size of the labeled training set, particularly for rare classes. Future work involves expanding training data to TESS extended missions, refining the model for TESS-Gaia light curves (TGLC), and exploring self-supervised learning to reduce reliance on labeled data.

In summary, ASTRAFier represents a significant advancement in automated stellar classification, demonstrating that hybrid deep learning architectures can effectively replace traditional feature-engineering pipelines for large-scale astronomical time-series analysis.