Applying multimodal learning to Classify transient Detections Early (AppleCiDEr) I: Data set, methods, and infrastructure
Abstract
Modern time-domain surveys like the Zwicky Transient Facility (ZTF) and the Legacy Survey of Space and Time (LSST) generate hundreds of thousands to millions of alerts, demanding automatic, unified classification of transients and variable stars for efficient follow-up. We present AppleCiDEr (Applying Multimodal Learning to Classify Transient Detections Early), a novel framework that integrates four key data modalities (photometry, image cutouts, metadata, and spectra) to overcome limitations of single-modality classification approaches. Our architecture introduces (i) two transformer encoders for photometry, (ii) a multimodal convolutional neural network (CNN) with domain-specialized metadata towers and Mixture-of-Experts fusion for combining metadata and images, and (iii) a CNN for spectra classification. Training on ~ 30,000 real ZTF alerts, AppleCiDEr achieves high accuracy, allowing early identification and suggesting follow-up for rare transient spectra. The system provides the first unified framework for both transient and variable star classification using real observational data, with seamless integration into brokering pipelines, demonstrating readiness for the LSST era.
Cite
@article{arxiv.2507.16088,
title = {Applying multimodal learning to Classify transient Detections Early (AppleCiDEr) I: Data set, methods, and infrastructure},
author = {Alexandra Junell and Argyro Sasli and Felipe Fontinele Nunes and Maojie Xu and Benny Border and Nabeel Rehemtulla and Mariia Rizhko and Yu-Jing Qin and Theophile Jegou Du Laz and Antoine Le Calloch and Sushant Sharma Chaudhary and Shaowei Wu and Jesper Sollerman and Niharika Sravan and Steven L. Groom and David Hale and Mansi M. Kasliwal and Josiah Purdum and Avery Wold and Matthew J. Graham and Michael W. Coughlin},
journal= {arXiv preprint arXiv:2507.16088},
year = {2025}
}
Comments
17 pages, 8 figures