Multimodal multi-output ordinal regression for discovering gravitationally-lensed transients

Gravitational lenses are caused by massive bodies that distort space-time, bendinglight. They can distort transients, such as Supernovae (SN), which are being studiedextensively. Gravitationally-lensed supernovae (LSN) are rare, so only a few havebeen discovered. Future astronomical surveys will collect huge amounts of data,calling for automated and accurate discovery techniques to find them. Still, onlya few works aim to discover LSN, most use only a few classes to characterizecandidate observations, and only a few exploit spatial and temporal information.This work introduces AstroCountNet (ACoNet), an ensemble of multimodal neuralnetworks that takes in input spatio-temporal data and, for each observation, countsthe occurrences of 7 astronomical bodies. ACoNet achieves, on average, morethan 85% macro F1 score on four datasets. The network is then adapted intoAstroClassNet (AClaNet) to address classification problems, achieving macro $F_1$scores between ≈ 59% and ≈ 93%.