🔭 astrophysics

EncoTESS: Age-Sensitive Encodings from Raw TESS Light Curves

The paper introduces EncoTESS, a compact Time Series Foundation Model trained on TESS light curves that effectively handles observational noise and irregular sampling to encode stellar variability into a latent space, demonstrating superior performance over traditional metrics in inferring the ages of young K and M stars.

Original authors: Phil R. Van-Lane (David A. Dunlap Department of Astronomy and Astrophysics, University of Toronto, Dunlap Institute for Astronomy and Astrophysics, University of Toronto, Department of Astronomy and A

Published 2026-08-27

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Phil R. Van-Lane (David A. Dunlap Department of Astronomy and Astrophysics, University of Toronto, Dunlap Institute for Astronomy and Astrophysics, University of Toronto, Department of Astronomy and Astrophysics, University of California San Diego), Joshua S. Speagle (David A. Dunlap Department of Astronomy and Astrophysics, University of Toronto, Department of Statistical Sciences, University of Toronto, Dunlap Institute for Astronomy and Astrophysics, University of Toronto, Data Sciences Institute, University of Toronto), Ryan Cloutier (Department of Physics and Astronomy, McMaster University), Christopher A. Theissen (Department of Astronomy and Astrophysics, University of California San Diego), Gwendolyn M. Eadie (David A. Dunlap Department of Astronomy and Astrophysics, University of Toronto, Department of Statistical Sciences, University of Toronto, Data Sciences Institute, University of Toronto), Ilay Kamai (Physics Department, Technion Israel Institute of Technology)

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

Stars are not the static, unchanging points of light they appear to be in the night sky. Even the steady glow of a sun-like star often flickers with a subtle, rhythmic pulse. This variation is caused by the star spinning, carrying dark, cool spots across its face like clouds drifting over a landscape, and by sudden, violent bursts of energy known as flares. For astronomers, these fluctuations are more than just visual noise; they are a clock. The speed at which a star spins and the intensity of its flares change predictably as the star ages. By measuring these patterns, scientists can estimate a star's age, a task that is otherwise incredibly difficult because stars of the same type can look nearly identical regardless of whether they are young or old.

However, reading this cosmic clock has traditionally been a difficult job. The patterns are often messy, obscured by gaps in observation, and complicated by the fact that different types of stars spin down at different rates. For the youngest and reddest stars, the old methods of timing their rotation often fail, leaving astronomers with a blurry picture of their history. To solve this, a team of researchers has developed a new approach that treats the entire light curve—the complete record of a star's brightness over time—as a single, complex story to be read all at once, rather than trying to extract a few isolated numbers from it.

The researchers, led by Phil Van-Lane and colleagues, created a specialized computer program called EncoTESS. This tool is designed to look at the raw data streaming from the Transiting Exoplanet Survey Satellite, or TESS, which watches the sky for tiny dips in starlight. Instead of trying to force the data into a simple, pre-defined shape, the program learns to recognize the unique "fingerprint" of variability in each star. It is trained on nearly 70,000 light curves from over 21,000 stars, learning to distinguish between the smooth rhythm of rotation, the chaotic spikes of flares, and the random static of instrumental noise. The program is remarkably efficient; it is small enough to run on a standard laptop, yet it captures the essence of the data in a way that traditional methods cannot.

When the team tested this new system, they found it could reconstruct the light curves of stars with high accuracy, filling in gaps and predicting future brightness changes better than simpler models. More importantly, the program learned to compress the complex, jagged lines of a star's brightness into a compact set of numbers that describe the star's physical state. These compressed descriptions, or encodings, contained hidden information about the star's age that was not obvious from its rotation speed alone. The system proved particularly powerful for young, red stars, which are notoriously difficult to date. In these cases, the new method outperformed the standard techniques that rely solely on measuring how fast a star spins.

The study reveals that the age of a star is written not just in its spin, but in the subtle details of how its light varies over time. By analyzing the entire history of a star's brightness, including the shape of its flares and the texture of its noise, the program can infer an age that aligns closely with known values for star clusters. This suggests that the chaotic, stochastic behavior of young stars holds a key to their history that was previously overlooked. While the method works best for stars that have not yet settled into a slow, steady rotation, it offers a promising new path for understanding the life cycles of the most common stars in our galaxy. The researchers have made their code and the resulting data available to the scientific community, opening the door for others to apply this technique to other surveys and to explore the hidden stories written in the light of distant suns.

Technical Summary: EncoTESS – Age-Sensitive Encodings from Raw TESS Light Curves

Problem Statement
Stellar age estimation for main sequence (MS) stars of spectral types late F through M remains a significant challenge. Traditional methods face specific limitations: isochrone fitting is ineffective for single MS stars due to minimal color-magnitude evolution; asteroseismology is difficult to apply to cooler FGKM stars due to low oscillation amplitudes; and lithium depletion becomes unreliable for stars older than $\sim1$ Gyr. While gyrochronology (using rotation period, $P_{\text{rot}}$ ) is a standard proxy, it suffers from intrinsic scatter and multimodality, particularly for young, red stars (K and M dwarfs $\lesssim 100$ Myr and M dwarfs $\lesssim 1$ Gyr) that have not yet converged onto the slow rotator sequence. Furthermore, existing light curve (LC) derived metrics (e.g., flaring rates, magnetic activity indices, variability amplitude) are often stochastic, dependent on observation time, or require hand-picked summary statistics that lose information and are difficult to optimize jointly due to correlations.

Methodology
The authors introduce EncoTESS, a Time Series Foundation Model (TSFM) specifically designed to encode raw TESS 2-minute cadence light curves into a fixed-size latent parameter space. Unlike task-specific models, EncoTESS is trained via self-supervised reconstruction to learn generalized patterns of time series data.

Architecture: EncoTESS utilizes a Recurrent Neural Network (RNN) based on the minGRU architecture (Feng et al. 2024), rather than the computationally expensive Transformer architectures common in other TSFMs. This choice allows the model to handle long sequences efficiently ( $O(L)$ scaling vs. $O(L^2)$ ) and run on consumer hardware with only $\approx 64,000$ parameters.
Input Data: The model ingests raw, normalized flux ( $\hat{f}$ ), measurement uncertainty ( $\hat{\sigma}_f$ ), and observation time ( $t$ ) as a time series. Crucially, it also incorporates a 13-dimensional metadata vector including stellar properties (Gaia $G_0$ , $(BP-RP)_0$ , parallax), TESS observational parameters (magnitude, sector, camera, CCD), and normalization statistics. Time is treated as an observable to handle irregular sampling and data gaps (e.g., TESS downlink times) without interpolation.
Training: The model is trained on 69,345 light curves from 21,507 unique stars using unlabeled self-supervised reconstruction. It predicts flux values at variable "horizons" (from 2 minutes to 4 days) using a Conditional Normalizing Flow (CNF) head to estimate full probabilistic likelihoods rather than point estimates. Masking is applied during training to ensure robustness against missing data.
Output: The RNN generates a sequence of hidden states ( $H_t$ ) representing the LC context at each time step. These are aggregated via pooling into a fixed 1536-dimensional latent vector ( $\Theta$ ). For age inference, a Partial Least Squares (PLS) projection creates a 3-dimensional subspace ( $\Theta^3_\tau$ ) optimized to maximize covariance with stellar age while remaining linearly orthogonal to rotation period.

Key Contributions

A Specialized TSFM for TESS: Unlike previous astronomy foundation models trained on Kepler data (which have different noise characteristics and baselines), EncoTESS is trained directly on noisier, shorter TESS light curves and explicitly handles heteroskedastic measurements and irregular sampling.
Efficiency: The model is $\approx 1\%$ the size of typical literature TSFMs, enabling execution on modern laptops.
Holistic Variability Encoding: EncoTESS encodes raw light curves without relying on pre-computed summary statistics, preserving information about rotation, flaring, and higher-order flux distribution moments.
Age Inference Framework: The authors demonstrate that the latent encodings can be used to infer stellar ages probabilistically, outperforming traditional gyrochronology for specific stellar populations.

Results

Flux Reconstruction: EncoTESS outperforms baseline models (windowed mean, metadata-only, and MLP-based models) in negative log likelihood (NLL), particularly at prediction horizons of hours to days. It effectively captures long-term trends and handles observational noise.
Variability Characterization: The latent space ( $\Theta$ ) successfully recovers traditional variability summary statistics (rotation period, flaring equivalent duration, skewness, kurtosis) with higher accuracy than models using only metadata or short-term context. The encodings cluster stars by physical properties (e.g., eclipsing binaries vs. flaring M dwarfs) and correlate strongly with color, magnitude, and variability amplitude.
Age Inference Performance: In a "leave-one-sector-out" cross-validation test, the EncoTESS-based age estimator ( $\text{age}_{\text{ET}}$ $age_{ET}$ ) outperforms a standard gyrochronology model ( $\text{age}_{\text{P}}$ $age_{P}$ ) and a model combining rotation and variability amplitude ( $\text{age}_{\text{PfH}}$ $age_{PfH}$ ) across most metrics (Pearson $r$ $r$ , Median Absolute Error, and bias).
- Specific Regime: The improvement is most pronounced for young, red stars (late K and M dwarfs) that have not converged onto the slow rotator sequence (K/M stars $\lesssim 100$ Myr; M stars $\lesssim 1$ Gyr).
- Field Stars: For older field stars (exoplanet hosts), the model's performance is comparable to models using only color-magnitude diagram (CMD) and variability metadata, suggesting the RNN architecture captures age information primarily in the younger, more active regime.

Significance and Claims
The paper claims that EncoTESS provides a data-driven, holistic approach to stellar characterization that unifies different types of variability (quasi-sinusoidal rotation and stochastic flaring) into a single representation. The authors assert that EncoTESS successfully extracts age-dependent variability signatures that are complementary to rotation period, offering a significant improvement in age constraints for the youngest and reddest stars where gyrochronology is least precise.

The work establishes that a lightweight, RNN-based foundation model trained on raw TESS data can effectively disentangle astrophysical signals from observational systematics. The authors conclude that this framework offers a promising foundation for future extensions to other cadences and surveys (Kepler, K2, PLATO) and for downstream tasks such as stellar classification and variability detection, while maintaining a library of encodings that can be updated as new data becomes available.

More like this