2026 workshop program
Fast and Flexible Unsupervised Characterization of Astronomical Time Series with Multi-Time Attention
The work
Abstract
As sky surveys grow in scale and cadence, the ability to rapidly and automatically characterize astronomical time series becomes essential. I present a simple and efficient method based on the Multi-Time Attention Network (mTAN) that learns time-aware latent representations of photometric light curves, with no imputation required for irregular or incomplete observations.
Using ZTF alert data, I show that the model produces accurate interpolations even for sparse light curves, organizes a latent space around physically meaningful properties (duration, peak time, variability, color), separates SN and AGN populations without supervision, and generalizes to unseen classes, including long-period variables and TDEs. Attention map visualizations reveal the model’s capability to capture local structure, demonstrating ML interpretability. The method is lightweight—it requires a few hundred kilobytes per model and is fast—the inference time is 0.01 seconds per light curve and scales as O(1) with observation count. This lightweight and scalable approach can be integrated into Rubin brokers to characterize time series at scale.