The coming era of the Vera C. Rubin Observatory’s Legacy Survey of Space and Time (LSST) will deliver an unprecedented torrent of alerts—millions of detections per night from a universe full of things that go bump in the dark. Meeting this challenge requires not just faster algorithms, but a fundamentally new approach to AI for astronomy.

As Deputy Director for Astrophysics at the NSF-Simons SkAI Institute, I co-lead the development of SELDON (Supernova Explosions Learned by Deep ODE Networks), a foundation AI model purpose-built for time-domain astrophysics, with Prof. Noelle Samia (Northwestern). SELDON processes multimodal inputs—light curves, spectra, and images from multiple surveys and multi-messenger experiments—handling the sparse, irregularly sampled data that characterize real astronomical observations. A key capability is predicting the future evolution of a supernova light curve even when only a handful of pre-peak detections are available. The initial SELDON paper appeared in the main technical track of AAAI-2026. A production version capable of processing the full LSST alert stream is in active development.

This work builds on a long program of AI for transient discovery and classification. I am a lead developer of the ANTARES alert broker, operating on the Zwicky Transient Facility (ZTF) alert stream as a testbed for LSST. I led the validation team for the Photometric LSST Astronomical Time-series Classification Challenge (PLAsTiCC), and co-lead the Extended PLAsTiCC (ELAsTiCC)—the largest simulation of the time-domain sky ever created and the largest Kaggle challenge in astronomy. I also lead the RESSPECT project, which applies active learning to intelligently prioritize spectroscopic follow-up resources for LSST, and the El-Cid filter for rapid identification of electromagnetic counterparts to gravitational wave events.

Complementing the real-time classification work, I am developing transformer models using transfer learning to classify supernova spectra, incorporating wavelet-based denoising to handle noise and instrumental variations in public spectroscopic datasets.

SCiMMA: Cyberinfrastructure for Multi-Messenger Astrophysics

As PI of SCiMMA (Scalable Cyberinfrastructure to support Multi-Messenger Astrophysics), I lead a collaboration of astronomers, computer scientists, and data scientists building shared community infrastructure for the multi-messenger era. SCiMMA develops and operates several interconnected services:

• Hopskotch — A publish-subscribe and archive platform built on Apache Kafka, enabling real-time distribution of alerts from gravitational wave, neutrino, gamma-ray, and optical facilities. Messages are retained for 30 days with optional long-term archival, and the platform supports binary, AVRO, VoEvent, and JSON formats.
• Blast — A real-time host galaxy characterization service that automatically identifies available archival data and measures star formation rates, stellar masses, and stellar ages for transient host galaxies reported to the IAU.
• DASH — Originally developed by Daniel Muthukrishna (Muthukrishna et al. 2019), DASH uses a deep learning model to classify supernova spectra. The SCiMMA service makes this available without requiring local installation, and also hosts an in-development transformer-based classifier from the SELDON team at the SkAI Institute that improves upon the original model.
• HEROIC — A coordination platform that displays real-time facility status and streamlines observing requests for multi-messenger follow-up campaigns.

SCiMMA also hosts OpenMMA, a community forum (successor to OpenLVEM) with regular community calls and a listserv for discussion of observatories, infrastructure, and multi-messenger science.