EncoTESS: Age-Sensitive Encodings from Raw TESS Light Curves
Authors: Phil R. Van-Lane, Joshua S. Speagle, Ryan Cloutier, Christopher A. Theissen, Gwendolyn M. Eadie, Ilay Kamai
Organizations: 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 · David A. Dunlap Department of Astronomy & Astrophysics, University of Toronto, 50 St. George St, Toronto, ON M5S 3H4, Canada · Dunlap Institute for Astronomy & Astrophysics, University of Toronto, 50 St. George St, Toronto, ON M5S 3H4, Canada · Department of Astronomy & Astrophysics, University of California San Diego, 3183 Matthews Ln, La Jolla, CA 92093, USA · Department of Statistical Sciences, University of Toronto · Data Sciences Institute, University of Toronto · Department of Statistical Sciences, University of Toronto, 9th Floor, 700 University Ave, Toronto, ON M5S 3G3, Canada · Data Sciences Institute, University of Toronto, 10th Floor, 700 University Ave, Toronto, ON M7A 2S4, Canada · Department of Physics and Astronomy, McMaster University · Department of Physics & Astronomy, McMaster University, 1280 Main St West, Hamilton, ON L8S 4L8, Canada · Physics Department, Technion Israel Institute of Technology · Physics Department, Technion: Israel Institute of Technology, Haifa 32000, Israel
Main sequence stars of spectral types late F through M exhibit systematic variability in photometric light curves, particularly when they are young. Rotational modulation of starspots manifests as quasi-sinusoidal variability, which enables the measurement of rotation periods. Variability can also be stochastic, as in stellar flaring. However, since measurements of stochastic processes depend on the time of observation, they are typically noisier. Considering that different manifestations of variability have unique observational nuances, models that naturally unify these are incredibly useful for stellar characterization. Towards this goal, we have developed EncoTESS: a Time Series Foundation Model (TSFM) trained on a subset of TESS 2-min light curves. EncoTESS is specifically designed to handle the observational noise, heteroskedastic measurements, irregular sampling, and large data gaps common to TESS data. It is also ~1% of the size of a typical literature TSFM, so can be run easily on a modern laptop. EncoTESS encodes light curves into a fixed-size latent parameter space, which can be used to infer physical stellar properties and recovers light curve summary statistics well. EncoTESS outperforms rotation period and variability amplitude as age indicators for stars that have not converged onto the slow rotator sequence yet; broadly these include K and M stars less than ~100 Myr, and M stars less than ~1 Gyr. We focus on age inference as an application of EncoTESS in this work, but other downstream tasks such as stellar classification could also be explored. The architecture of EncoTESS enables its future extension to TESS light curves of all cadences, and additional surveys such as Kepler and the upcoming PLATO mission. The core EncoTESS framework and library of encodings produced for the stars used in this work are publicly available at https://github.com/philvanlane/encotess.
Figures & tables
Reference
# Stars
# LCs
Van-Lane et al. [104]
2,495
7,166
Feinstein et al. [27]
6,753
25,198
Kiman et al. [51]
753
2,964
Engle [21]
54
205
Magaudda et al. [61]
154
372
Mamonova et al. [62]
320
1,985
Table 1: Number of stars and total LCs ( TESS sectors) in the EncoTESS training catalog, broken down by literature source.
Figure 1: Distribution of Gaia DR3 (BP−RP)0 colors by source category. Exoplanet hosts and thick disk stars are highly concentrated around (BP−RP)0≈1 , whereas the other targets are more evenly distributed are are responsible for the other peak at (BP−RP)0≈3 .
Figure 2: A schematic view of the data flow and training process of EncoTESS . The training dataset consists of raw LCs (timestamp, flux, and flux measurement uncertainty values for each point), and their corresponding metadata consisting of stellar and sector parameters from Gaia DR3 and TESS . Training is via reconstruction: subsets of an LC are encoded using both forward and backward processing into hidden states ( Htf and Htb respectively), which are used to predict the flux at some offset. EncoTESS learns to construct these hidden states to maximize the likelihood of its predictions. This is repeated at variable prediction offsets. After training is complete, the hidden states corresponding to every observation in a LC are aggregated into a 1536-dim latent representation of each LC. In this schematic, solid lines represent one-time data processing and dashed lines represent computations that are executed iteratively during training. The colored grids represent vectors, where each pixel is a dimension and the color represents the magnitude of that dimension. The dimensions shown here are accurate: each Htf and Htb is 64-dim, so each Ht is a 128-dim concatenation of Htf and Htb , and the final LC encoding is a 1536-dim vector constructed by pooling all Ht vectors in 12 different ways (see Appendix C.4 for details).
Figure 3: The confidence of flux predictions is highly dependent on the prediction horizon. Here we compare forward model predictions where the horizon is (i) 0.5 hours, and (ii) 17 hours. Shaded regions represent the 1 σ confidence interval of predictions. Black points with the grey error bars are true measured values, normalized as per Equation 1 .
Figure 4: The confidence of flux predictions also depends on the measurement error. Here we compare flux predictions using the real measurement error against reconstructions based on models with artificially dampened (0.3 × true) and inflated (3 × true) TESS flux uncertainty. Shaded regions represent the 1 σ confidence interval of p(f^) in each scenario. Black points with the grey error bars are true measured values, normalized as per Equation 1 . Error bars are not shown in the prediction regime to avoid clutter. While the credible regions do not always symmetrically widen at every point with inflated σ^f , the uncertainty increases overall.
Figure 5: Flux reconstruction performance for the four model variants, using negative log likelihood (NLL) and median absolute errors (MAE) as metrics. Individual NLL values were clipped to ≈10σ for stability. The RNN architecture and CNF prediction head used in EncoTESS provide value, particularly when looking at prediction horizons of a ∼ few hours to days and when considering the entire likelihood shape, although the overall accuracy of EncoTESS and the MLP are comparable.
Figure 6: LC encodings ( Θ ) for two sample stars. The bottom right panel shows each star’s position on a color-magnitude diagram. The bottom left panel shows the positions of each Θ , compressed into two dimensions with UMAP . The background points are all other stars in our catalog. Example LCs from each star are shown in the top panel, which correspond to the circled LCs in the UMAP plot. The median distance between points ( dmed ) representing different LCs from the same star is plotted for reference.
Figure 7: All LC encodings, compressed from the 1536-dim Θ into 2 dimensions with UMAP . Points are color coded by the Gaia (BP−RP)0 color of the star.
Figure 8: All stars in our dataset, positioned by two dimensional UMAP compression of their latent encodings. Each subplot shows the stars colour-coded by specific measured parameters, which are clearly tractable in all cases through the latent representations.
Figure 9: Violin plot comparing the recovery of 4 LC summary statistics from 4 different inputs: (i) The real Θ computed by EncoTESS , (ii) pure noise, (iii) only metadata, and (iv) an MLP-derived baseline. (iii) and (iv) are analogous to the metadata and MLP variants described in § IV.2 , but with a metric recovery target instead of flux reconstruction target. EncoTESS performs best overall, indicating that our architectural choices provide value. Residuals are transformed and scaled for ease of visualization, but key metrics are provided in native units in Table 2
Model
skew( fn )
kurt( fn )
Prot (d)
ED flare (min)
Med.
IQR
Med.
IQR
Med.
IQR
Med.
IQR
EncoTESS
0.002
0.037
0.013
0.091
0.01
0.67
0.00
3.22
MLP
0.014
0.084
0.101
0.337
0.03
0.71
0.00
3.62
metadata
0.027
0.310
0.273
0.847
-0.09
2.20
-0.01
5.10
noise
0.134
0.181
0.675
0.986
-0.04
5.33
-0.66
8.51
Table 2: Median and interquartile range (IQR) of the residuals for each model/metric combination from a recovery test, in native units.
Figure 10: Colour coded contribution of every point in the LC (TIC 280865190, sector 7) to PC1, measured using an integrated gradient (IG) technique and occlusion. Positive and negative contributions move PC1 in different directions, but the magnitude of the contribution can be considered as an “importance” score.
Figure 11: Analogous to 10 , but for the eclipsing binary TIC 394508465, sector 39.
Figure 12: Saliency analysis of the eclipsing binary TIC 394508465, sector 39, comparing the magnitudes of the integrated gradients of the forward and backward hidden states. It is clear that the forward state is more sensitive to individual measurements overall, but the backward state is particularly sensitive to the eclipse features.
Figure 13: Top 10 light curve features that have the strongest correlation with the PC1 of Θ′ . We included metadata features in this analysis as well as age τ (as provided in the source catalogs) and the variability statistics shown in Figure 8 . It is clear that PC1 is primarily driven by the overall level of variability and target brightness, and color/magnitude measurements.
Figure 14: Schematic showing age inference for two stars that are in similar places on a color-magnitude and color-rotation diagram, but have different light curve properties and different encodings. ageP is unable to recover accurate ages from rotation alone, but ageET is able to recover very accurate ages using the light curve encodings.
Figure 15: Features other than Prot that are most aligned with age in the EncoTESS latent space. In the left panel we measure the correlation of each feature with age in the 3-dim Θ3τ space that is linearly orthogonal to Prot . We do the same in the right panel, but in a space linearly orthogonal to the primary Prot / MG / (BP−RP)0 direction. In each panel we only show the top six features; the correlation with any other features is r≤0.02 .
Figure 16: All LC encodings, compressed from 1536 dimensions into 2 dimensions with UMAP . Points are colour coded by the age of the star from the literature catalogs described in § II , and clear structure is visible.
Method
Pearson r
MAE
median res
% ∣ res ∣
Coverage
(dex)
(dex)
>0.5 dex
p16-p84 (%)
All Stars
ageET ( Θ3τ )
0.674
0.431
+ 0.028
31.0
72.1
agePfH ( Prot,fH )
0.601
0.496
− 0.039
34.7
66.5
ageP ( Prot )
0.532
0.529
− 0.064
38.7
72.8
KM Subset
Table 3: Results from three different age recovery models: ageET which uses the 3-dim Θ3τ as its indicator, agePfH which uses both Prot and fH as indicators, and ageP which uses only Prot as its age indicator. The models are compared using 5 different metrics: Pearson r , median absolute error (MAE), median residual, % of residuals >0.5 dex, and % of true ages covered by the p16–p84 posterior interval. This last metric should be close to 68% for a well calibrated model. In 4/5 metrics, ageET outperforms agePfH which outperforms ageP . In p16-p84 coverage, agePfH outperforms the other two models, which both exhibit overly conservative posteriors (although only by 4–7%). We have measured these metrics over all stars in our catalog (upper block), and again over only the region of relatively young, active stars (we call this the KM region for simplicity) as outlined in Figures 18 and 19 . The difference between ageET and the other 2 models is larger in this regime.
Figure 17: A comparison of ages inferred from ageET and ageP . The left panel plots the residuals against (BP−RP)0 , with approximate spectral types shown on top. The middle panel plots them against true age; we place stars along the x-axis evenly here in order of increasing age for ease of interpretation. The right panel shows the overall distribution of residuals. ageET is more precise for younger and redder stars, and less biased to the mean than ageP . From the histogram we also see that ageET is less systematically biased and has a narrower distribution of residuals than ageP , indicating that its posterior age estimates are better calibrated.
Figure 18: Difference in LL values from ageET and ageP for stars in our catalog with measured rotation periods and ages. Blue regions indicate where ageET is stronger, and gold regions indicate where ageP is stronger by this metric. LL values are binned with a resolution of 0.01 in log(τinf) and 0.01 in (BP−RP)0 . Bins are smoothed over 0.2 in each direction to remove noise, and only bins with ≥5 stars are kept. Approximate spectral types are shown on top, as described in 17 . To the bottom right of the red dashed line is the KM region described in Table 3 , where ageET consistently outperforms ageP .
Figure 19: Analogous to Figure 18 but comparing ageET to agePfH . agePfH is stronger than ageP , but ageET still outperforms agePfH over the KM region.
Figure 20: Violin plot showing age recovery for the 2,430 exoplanet hosts in our catalog with ages in the default NASA Exoplanet Archive records, grouped into 2 Gyr age bins. EncoTESS performance is shown in blue, the noise baseline in grey, and the CMD-only and CMD+variability variants in gold and green respectively. The majority of ageETXH ’s predictive power for these stars comes from CMD indicators, with high level variability contributing in the 0-2 Gyr bin. The number of stars in each bin is shown on top of the plot. Past ≈6 Gyr, all models struggle.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 21: Reconstruction of the power spectrum of TIC 67787319 Sector 43. The left panel shows the reconstruction based on the core EncoTESS model, and the right panel shows the reconstruction when PS and ACF channels are included in training. Recovery of some features is marginally improved by including PS and ACF in training, but the overall shape is very similar.
Figure 22: Comparison of age (left) and Prot (right) inference tests using our Θ , compared between the final EncoTESS model and a version that included the PS and ACF encoders in training.
Figure 23: Age gradient over the top 8 PCs of Θ′ . It is apparent that the top component (PC1) carries strong age signal.
The discovery of exoplanets at scale has become one of the defining data science challenges in modern astrophysics. NASA's Transiting Exoplanet Survey Satellite (TESS) had catalogued over 7,800 planet candidates by early 2026, yet confirmation stands at fewer than 720. This paper introduces ExoNet, a multimodal deep learning framework that jointly processes phase-folded global and local light curve views alongside stellar parameter features through a calibrated late-fusion architecture combining 1D Convolutional Neural Networks, 8-head Multi-Head Attention over temporal feature maps, and a residual fusion head with post-hoc Temperature Scaling calibration. Trained on 7,585 labeled Kepler Objects of Interest, ExoNet achieves Test AUC = 0.9549 and 86.3% accuracy. Applied to 4,720 verified unconfirmed TESS Planet Candidates with TOI-TIC cross-identification verified against the NASA Exoplanet Archive, the model yields 1,754 high-confidence signals, 52 habitable-zone candidates, and six Earth-sized habitable-zone targets below 1.6 Earth radii. TOI-5728.01 and TOI-6716.01 emerge as the most Earth-like unconfirmed candidates. Full ablation confirms each modality improves AUC. Code and catalog are openly released.
Md. Rashadul Islam
Dept. of Computer Science and Engineering Daffodil International University Dhaka, Bangladesh
We present an end-to-end pipeline for estimating stellar parameters from Sloan Digital Sky Survey Data Release 12 spectra using a fully connected multitask neural network with residual blocks, whose hyperparameters are tuned via Bayesian optimization. The preprocessing pipeline includes per-spectrum standardization, RobustScaler normalization of the target variables -- effective temperature Teff, metallicity [Fe/H], and surface gravity logg -- and data augmentation via Gaussian noise injection. On a held-out test set, the model achieved Mean Absolute Errors (MAE) of 59.76K for Teff, 0.103dex for [Fe/H], and 0.130dex for logg. Normalized against the full-scale range of each parameter, these results represent range-normalized errors between 1% and 3%, achieved with a highly efficient model complexity of approximately 540,000 trainable parameters. These results demonstrate that a compact residual multitask architecture, combined with principled signal preprocessing, provides a parameter-efficient solution for nonlinear parameter estimation in large-scale spectral datasets. In particular, the proposed model achieves competitive performance with substantially lower complexity than deeper neural network baselines.
Bruno Santos Meneses Barreto, Marcio Eisencraft
Escola Polit´ecnica, Universidade de S˜ao Paulo, SP
Microlensing can reveal populations of faint compact objects that are otherwise difficult to detect. Depending on their design, all-sky surveys have the potential to search for these objects across the sky. The Transiting Exoplanet Survey Satellite (TESS), primarily designed to detect transiting exoplanets, also provides near all-sky coverage with high cadence. In this work, we use TESS data to search for microlensing candidates using both traditional and machine-learning methods and to identify associated false positives in high-cadence surveys. Microlensify is a physics-informed, transformer-based variational autoencoder trained on simulated single-lens microlensing light curves and real TESS Sector 12 data. The model classifies events, reconstructs light curves, and estimates microlensing event durations. Applied to ∼5.6 million TESS light curves, it identified between 0.036% and 1.89% as microlensing candidates across different TESS pipelines. After applying microlensing detection metrics and cross-matching with SIMBAD, we obtained a final list of candidates and identified false positives including long-period variables, Mira variables, cataclysmic variables, red giants, and transients. We also found Gaussian-like peaks caused by asteroid crossings, a potential source of false positives in high-cadence microlensing surveys. The model also predicts event duration with an accuracy of R2=0.97. The model was further tested on published events from different ground-based microlensing surveys, confirming 92.7% as microlensing, demonstrating its applicability across surveys with different cadences.
Department of Physics, Institute for Advanced Studies in Basic Sciences (IASBS), Zanjan 45137-66731, Iran · Department of Physics, Lehigh University, 16 Memorial Drive East, Bethlehem, PA 18015, USA · Department of Astronomy, University of Virginia, 530 McCormick Road, Charlottesville, VA 22904, USA +3