The authors present kilonova-multimodal-emulator, a simulation pipeline that makes realistic mock observations of kilonovae. Kilonovae are brief flashes of light that can follow the merger of two neutron stars or a neutron star and a black hole. Because real examples are rare, the team builds synthetic data — light curves (photometry), spectra, and telescope images — so large artificial‑intelligence models can be trained to find and study these events in the flood of alerts from sky surveys.

Kilonovae are hard to detect. They are short lived and sometimes faint. Wide‑field surveys such as the Zwicky Transient Facility (ZTF) and the Vera Rubin Observatory’s Legacy Survey of Space and Time (LSST) will produce millions of transient alerts each night. That volume makes automated tools essential. The simulated dataset is meant to teach AI not just to flag likely kilonovae but also to estimate physical parameters of the source.

To make realistic signals, the team uses physics‑based radiative‑transfer models. They generate a grid of model light curves and spectra using two radiative‑transfer families: POSSIS, a three‑dimensional Monte Carlo code, and the LANL TP2 family computed with the SuperNu code. From those model outputs they train fast machine‑learning surrogates. The surrogate or “emulator” combines a coordinate‑conditioned nonlinear predictor, a neural operator that works across wavelength, and a reduced‑order correction to capture remaining differences. Images are simulated to match ZTF instrument properties, including a realistic point‑spread function (PSF). For the demonstration they used historical cadence and limiting‑magnitude information from ZTF’s Bright Transient Survey (BTS).

This work matters because current catalogs do not contain enough real kilonova examples to train the large AI systems that will be needed. Existing public datasets like PLAsTiCC and ELAsTiCC focused on many classes of transients and on photometry alone. By contrast, kilonova‑multimodal‑emulator produces multi‑modal examples (photometry, spectra, and images) and targets one source class. That makes it a useful resource for training models aimed at identification and parameter estimation of kilonovae in real survey data.