Machine Learning for Time-Domain Astrophysics

Speaker: Alex Gagliano (IAIFI/MIT)

Venue&Time: Blue Room / 3:00 PM

Abstract: The time-evolving night sky is rich with variable stars, supernovae, and merging neutron stars. Wide-field imaging surveys that monitor this variability produce gappy, multi-modal observations that demand scalable, uncertainty-aware models for physical inference. In this talk, I’ll survey my recent work in building machine learning methods for time-domain astrophysics, with a focus on learning representations of our data for classification, physical inference and the discovery of astrophysical anomalies. I’ll introduce Minuet, a compact host-galaxy image encoder trained with diffusion modeling; and a mixture-of-experts model that fuses supernova light curves and spectra while preserving modality-specific information and yielding calibrated posteriors. I’ll conclude by outlining three areas at this intersection with the greatest potential to drive discovery in the coming years: better physical models, scalable population studies, and ML-guided survey optimization.