MAGIKS: Multimodal Alignment and Gravitational-Wave-Conditioned Interaction for Kilonova Search

Abstract

We present MAGIKS, a multimodal framework for gravitational-wave (GW)-triggered kilonova identification that formulates counterpart search as query-conditioned gallery retrieval. Following an align-before-fuse design, MAGIKS combines GW source parameters and sky-localization maps with irregularly sampled multiband optical light curves and sky coordinates. A time-aware supervised contrastive objective provides coarse cross-modal alignment, while dual cross-attention evaluates consistency between optical coordinates and GW localization, and between GW parameters and optical evolution. We train and evaluate MAGIKS on simulated binary neutron star and neutron star-black hole mergers, Rubin Observatory kilonova light curves from regular survey observations and target-of-opportunity follow-up, GW-detected mergers without Rubin-detectable kilonova counterparts, and realistic non-kilonova contaminants. In independent galleries with one true counterpart and time- and sky-compatible distractors, MAGIKS outperforms optical-only and sky-localization-only ranking, with increasing gains as galleries grow. At 5000 candidates, it ranks the counterpart first for 39.8% of queries, versus 4.8% and 0.2% for these baselines. In a GW170817 study, we generate counterfactual Rubin observations using released GW parameters and three-dimensional localization, the measured counterpart position, and host distance. At the same gallery size, MAGIKS achieves top-1 and top-10 retrieval rates of 87.3% and 95.7%. It correctly classifies 94.2% of associated and non-associated GW-optical pairs. In a separate crossed-pair experiment, its matching score responds significantly to controlled changes in all five tested GW parameters. These results support GW-conditioned kilonova candidate prioritisation for future multi-messenger follow-up.