HOLISMOKES

Strong gravitationally lensed supernovae (LSNe), though rare, are exceptionally valuable probes for cosmology and astrophysics. Upcoming time-domain surveys such as the Vera Rubin Observatory’s Legacy Survey of Space and Time (LSST) offer a major opportunity to discover large number of LSNe. Early identification is crucial for timely follow-up observations. We have developed a deep learning pipeline to detect LSNe using multiband, multi-epoch image cutouts. Our model is based on a 2D convolutional long short-term memory (<ns0:monospace>ConvLSTM2D</ns0:monospace>) architecture designed to capture both spatial and temporal correlations in time-series imaging data. Predictions are made after each observation in the time series, with accuracy expected to improve progressively as additional data are processed. We trained the model on realistic simulations derived from Hyper Suprime-Cam (HSC) data, which closely matches LSST in depth and filter characteristics. In this work, we focus exclusively on Type Ia supernovae (SNe Ia). LSNe Ia were injected into HSC luminous red galaxies (LRGs) at various phases of evolution to create positive examples of LSNe Ia time series. Negative examples include variable sources observed in the HSC Transient Survey (including unclassified transients) and simulated unlensed SNe Ia in LRG and spiral galaxies. Our multiband model shows rapid classification performance improvements during the initial few observations and quickly reaches a high detection efficiency: At a fixed false-positive rate (FPR) of 0.01%, the true-positive rate (TPR) reaches ≳60% by the seventh observation and exceeds ≳70% by the ninth observation. If we relax the FPR to 0.1%, the TPR reaches close to 60% as early as the fourth observation. Although the single-band analysis performs reasonably well in isolation, the multiband model significantly outperforms it, particularly in the early stages, by building a richer memory and leveraging color information. Among the negative examples, SNe in LRGs remain the primary source of FPR, as they can resemble their lensed counterparts under certain conditions. Additionally, the model detects quads more effectively than doubles, and it performs better on systems with larger image separations. Although we trained and tested the model on HSC-like data, our approach is applicable to any cadenced imaging survey – particularly LSST, where the higher expected cadence (five to ten times that of HSC) should further boost performance.