Astra-CLR: How Contrastive Learning Deciphers Multi-Year Light Curves ✛ ✎ ✦
Context
The big-data revolution in astronomy: modern surveys like ZTF and Gaia spew out millions of light curves nightly—records of brightness changes from galactic nuclei to individual stars. Manual classification is impossible, and spectroscopy, which provides physical parameters, is available for only a tiny fraction. To avoid drowning in this flood, astronomers need algorithms that can extract information directly from raw photometric series. That's where machine learning steps in, specifically self-supervised methods that don't require labeled data.
Methods
At its core, Astra-CLR uses a Transformer architecture adapted for irregular time series. The key innovation is multi-scale temporal 'views': from a single light curve, a long 'global' segment and several short 'local' ones are generated, each possibly augmented with noise, synthetic outliers, or masking. The network learns to match them through the NT-Xent contrastive loss, forcing it to learn a robust 'local-to-global' mapping. The data include three ZTF filters (g, r, i), and color information is introduced via separate effective wavelength embeddings, allowing the model to implicitly capture color evolution. During fine-tuning, late fusion is applied—aggregating representations from three non-overlapping windows (start, middle, end) covering the entire observation span.
Results
After pretraining on ∼2.1 million ZTF objects, the frozen model's performance was evaluated on a balanced sample of 120,000 light curves across 12 variability classes, spanning different stages of stellar evolution. These include Cepheids, RR Lyrae, eclipsing binaries, and cataclysmic variables with white dwarfs. Simple logistic regression on extracted vectors achieved 70.34% accuracy, while k-nearest neighbors reached 62.32%. But the real breakthrough came after 'label-efficient' fine-tuning: just 2% labeled data (∼42,000 curves) and unfreezing the Transformer's top layers boosted linear accuracy to 77.23% and k-NN to 76.42%. This proves that fine-tuning doesn't just shift hyperplanes—it radically improves the local structure of the latent space, with objects of the same class pulling into tight clusters, clearly visible in UMAP projections.
Implications
Astra-CLR is the first open-source multi-filter transformer trained solely on real ZTF data, without any metadata or spectra. It demonstrates that contrastive self-supervised learning can capture the subtle morphological features of variable stars, even under extremely irregular observing cadences. The learned representations serve as a ready-made foundation for classification, clustering, and anomaly detection in upcoming surveys like LSST and Roman, where manual labeling will become an unaffordable luxury.
Future development
Future development of the method envisions a shift to fully dynamic positional encoding that unifies time and color, better capturing the physics of multi-band observations. Another direction is cross-survey transfer learning—testing how well representations learned on ZTF apply to TESS or OGLE data. This is critical for building a universal 'foundation' astronomical embedder. Finally, based on the clustered latent space, a real-time system for automatic detection of rare transients is planned.
Impact
The developed approach directly impacts the automation of data analysis in large-survey astronomy and will accelerate scientific exploitation of LSST and Roman.
Next steps
Implementation of a color-aware dynamic positional encoder and large-scale testing of model transferability across surveys.
Key open problems
This tackles the problem of automatically classifying millions of variable stars without spectral data and the task of building universal representations for irregularly sampled time series.
The knowledge machine's view
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
The authors introduce Astra-CLR, a contrastive transformer trained on ~2.1 million unlabeled ZTF light curves. The model uses multi-scale representations (global and local windows) and three photometric filters. After fine-tuning on 2% of labeled data, it achieves classification accuracies of 77.23% (logistic regression) and 76.42% (k-NN) across 12 classes of variable stars.
Where its strength is
A key strength is its ability to extract informative representations from irregular multi-channel time series without external metadata, as demonstrated by both high accuracy and improved cluster structure in the latent space.
Where it could go next
- Use weak supervision for anomaly and rare-class detection: a multi-task learning approach with hierarchical labels from the gamma-ray sky analysis could be adapted to find unusual variable stars in Astra-CLR representations.based on: Gamma Sky as an Orchestra: Weak Supervision Catches Anomalies arXiv:2607.07158
- Apply augmentation techniques and architectural ideas from the stellar music work (e.g., converting light curves to spectrograms or using attention to periodicities) to further improve representations.based on: Melody, Harmony, Rhythm: How AI Learns to Listen to Stellar Music arXiv:2606.28446
Worth trying
- Train a classifier for rare classes (e.g., cataclysmic variables) using PU-learning or weak labeling, as done in the gamma-ray anomaly work.
- Test whether adding spectral features (derived from light curves via Fourier transform) improves clustering quality and downstream task accuracy.
Why this may matter
Advancing self-supervised learning on unlabeled photometric data is critical for scaling classification in the LSST era, where tens of millions of new variable objects are expected, while spectroscopy will remain limited.
Nearby works in our archive
- Teaching Neural Networks to Read Molecules at a Glancesimilarity 0.63
- Gamma Sky as an Orchestra: Weak Supervision Catches Anomaliessimilarity 0.62
- Why quasars flicker differently than we thoughtsimilarity 0.61
- Neural Network Discovers New Galaxies Without Any Guidancesimilarity 0.61
- Melody of Disorder: What a Black Hole's Light Says About Chaossimilarity 0.60
- Light Reveals the Secrets of a Layered Crystalsimilarity 0.60
- Melody, Harmony, Rhythm: How AI Learns to Listen to Stellar Musicsimilarity 0.60
The closest work in the archive sits at 0.63 by meaning — the usual level for our corpus.
The ZTF survey is named after Fritz Zwicky — the astrophysicist who not only predicted neutron stars and supernovae but also first proposed dark matter, and had a reputation as an eccentric genius who called his colleagues 'spherical bastards'.
Key numbers
- pretrained accuracy (linear): ~0.70
- fine-tuned accuracy (linear): ~0.77
- number of light curves for pretraining: ~2.1 million
- number of variability classes: 12
- number of filters: 3 (g, r, i)