Fast and Flexible Characterisation of Astronomical Light Curves Using Multi-Time Attention
Fuente:
arXiv
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Preprint |
| Published: |
2026
|
| Subjects: | |
| Online Access: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866910250190241792 |
|---|---|
| author | Gondhalekar, Yash Möller, Anais Sánchez-Sáez, Paula |
| author_facet | Gondhalekar, Yash Möller, Anais Sánchez-Sáez, Paula |
| contents | We present an unsupervised, data-driven framework for rapid characterisation of astronomical photometric time series using a Multi-Time Attention Network. The model learns time-aware latent representations directly from irregular, partial light curves without heavy preprocessing. Through application on ZTF alert data retrieved with Fink, a community alert broker for Rubin LSST, we demonstrate that the model: (i) produces accurate interpolations with small bias (0.01 mag) and scatter (0.1 mag) even for sparse light curves, (ii) learns a temporally distributed latent space correlating with physically meaningful properties (duration, peak time, variability, color) while being robust to unimportant properties such as observed magnitude and number of observations, (iii) separates general SN and AGN samples despite data being heavily dominated by AGNs, and (iv) generalises to unseen classes: The long-period variable and TDE show good interpolation and sensible latent space placement; however, the model cannot capture RRLyrae's $\sim$0.4-0.5 day pulsation period, which is far below our model's chosen two-day temporal resolution. Attention map analysis reveals the capability of multi-time attention to capture local structure. The model is extremely lightweight (a few hundred kilobytes) and has fast inference ($\sim$0.01 and $\sim$$3\times10^{-4}$ s per light curve on CPU and GPU, respectively) that is independent of the number of observations, unlike GP regression. Our approach offers flexible and scalable characterisation, with high relevance in the Rubin LSST era. We discuss future possibilities to incorporate observational uncertainties and symmetries for robustness and forecasting applications for real-time follow-up. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2605_24095 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Fast and Flexible Characterisation of Astronomical Light Curves Using Multi-Time Attention Gondhalekar, Yash Möller, Anais Sánchez-Sáez, Paula Instrumentation and Methods for Astrophysics Astrophysics of Galaxies We present an unsupervised, data-driven framework for rapid characterisation of astronomical photometric time series using a Multi-Time Attention Network. The model learns time-aware latent representations directly from irregular, partial light curves without heavy preprocessing. Through application on ZTF alert data retrieved with Fink, a community alert broker for Rubin LSST, we demonstrate that the model: (i) produces accurate interpolations with small bias (0.01 mag) and scatter (0.1 mag) even for sparse light curves, (ii) learns a temporally distributed latent space correlating with physically meaningful properties (duration, peak time, variability, color) while being robust to unimportant properties such as observed magnitude and number of observations, (iii) separates general SN and AGN samples despite data being heavily dominated by AGNs, and (iv) generalises to unseen classes: The long-period variable and TDE show good interpolation and sensible latent space placement; however, the model cannot capture RRLyrae's $\sim$0.4-0.5 day pulsation period, which is far below our model's chosen two-day temporal resolution. Attention map analysis reveals the capability of multi-time attention to capture local structure. The model is extremely lightweight (a few hundred kilobytes) and has fast inference ($\sim$0.01 and $\sim$$3\times10^{-4}$ s per light curve on CPU and GPU, respectively) that is independent of the number of observations, unlike GP regression. Our approach offers flexible and scalable characterisation, with high relevance in the Rubin LSST era. We discuss future possibilities to incorporate observational uncertainties and symmetries for robustness and forecasting applications for real-time follow-up. |
| title | Fast and Flexible Characterisation of Astronomical Light Curves Using Multi-Time Attention |
| topic | Instrumentation and Methods for Astrophysics Astrophysics of Galaxies |
| url | https://arxiv.org/abs/2605.24095 |