No abstract

We present ATAT, the Astronomical Transformer for time series And Tabular data, a classification model that receives as input both light-curves and tabular data from astronomical sources. ATAT consists of a light-curve Transformer with a new time modulation that encodes the time of each observation, and a feature Transformer that uses a Quantile Feature Tokenizer. This model was conceived by the ALeRCE alert broker in the context of the recent Extended LSST Astronomical Time-Series Classification Challenge (ELAsTiCC). ATAT outperforms previous decision tree-based ensemble approaches in terms of classification when trained over the ELAsTiCC dataset. Importantly, some of its variants do not require human-engineered features, with significantly reduced inference computational times (400x faster). The use of Transformer multimodal architectures, combining light-curve and tabular data, opens new possibilities for classifying alerts from a new generation of large etendue telescopes, such as the Vera C. Rubin Observatory, in real-world brokering scenarios.