For centuries, astronomers gazed at the night sky, catching the slightest changes in stellar brightness. These uneven sequences, light curves, are like musical staves: each star performs its own unique melody. But today, data streams from telescopes like the Vera Rubin Observatory (LSST) have turned into a deafening cacophony — billions of 'recordings' nightly. Manual deciphering of these cosmic scores, based on features from the time of Cecilia Payne-Gaposchkin, has become impossible. Machine learning comes to the rescue, but how do you teach an algorithm to understand the language of pulsating stars?

The new work proposes an elegant solution: instead of a single viewpoint, the model views the star through three 'lenses'. The first view is the raw time series: uneven observations create a jagged rhythm, like a melody with missing bars. The second is the periodogram (Lomb–Scargle spectrum): it decomposes chaos into harmonics, like a chord revealing the dominant frequency. The third is the phase curve, folded by the best period: a pure rhythmic pattern, the shape of the star's own 'beat'. The LeJEPA architecture does not analyze them separately: it forces the three projections to 'sound in unison'. Unsupervised self-distillation aligns the latent representations of each view, and the model captures the invariant essence of the star, ignoring noise. Two key innovations embed physical intuition directly into the model's 'hearing'. Continuous Rotational Positional Encoding (C-RoPE) marks the observation time not by position in the array, but by the actual astronomical date. Imagine: note durations are recorded not by measures from the score's start, but by absolute time: this makes the representation translation-invariant. And Error-Aware Numeric Embedding (EANE) treats the measured brightness not as a point, but as a fuzzy Gaussian cloud. If the telescope reports 'the star shines at 15.2, but the uncertainty is large,' the model averages embeddings over a probable range, and noise stops deceiving. Thus, something like a musical ear is born: the algorithm learns to hear the cosmos much like an infant absorbs its native language — without preset rules, by immersing in patterns.

The results are staggering. On the StarEmbed benchmark (tens of thousands of real light curves from the Catalina Surveys), the multi-view model outperformed meticulously engineered expert features on 15 out of 16 metrics — regardless of the classifier. The real triumph lies in the minimal-label scenario. Just one example per variable star class yields a macro-F1 of 42.5% versus 30.3% for the manual approach. With 100 examples, accuracy reaches 63.6%, confidently beating the classic. This proves the algorithm truly learned a physically meaningful representation, not just memorized data. Visualizations of the latent space show how object classes (from Cepheids to supernovae, products of stellar evolution) self-organize into compact islands, even though the model never saw labels.

This work is not just an improvement in classification. It marks a shift from handcrafted feature engineering to automatic construction of physically motivated encoders. In the future, integrating a differentiable periodogram and learnable phase folding could eliminate the manual stage entirely — the model itself would discover hidden periodicities. And by combining photometry with spectroscopy from JWST-class telescopes, we will learn to hear the 'timbre' of a star, its chemical composition and temperature, and even notice tiny dips from exoplanets. Such tools directly impact the accuracy of cosmological distances via standard candles and redshift, reducing tension in the Hubble constant and clarifying the nature of dark energy. This opens the path to statistical studies of star formation and galactic evolution, and eventually to solving inverse problems of numerical simulation on big-data scales.