ParSNIP: Generative Models of Transient Light Curves with Physics-enabled Deep Learning

We present a novel method to produce empirical generative models of all kinds of astronomical transients from data sets of unlabeled light curves. Our hybrid model, which we call ParSNIP, uses a neural network to model the unknown intrinsic diversity of different transients and an explicit physics-based model of how light from the transient propagates through the universe and is observed. The ParSNIP model predicts the time-varying spectra of transients despite only being trained on photometric observations. With a three-dimensional intrinsic model, we are able to fit out-of-sample multiband light curves of many different kinds of transients with model uncertainties of 0.04–0.06 mag. The representation learned by the ParSNIP model is invariant to redshift, so it can be used to perform photometric classification of transients even with heavily biased training sets. Our classification techniques significantly outperform state-of-the-art methods on both simulated (PLAsTiCC) and real (PS1) data sets with 2.3× and 2× less contamination, respectively, for classification of Type Ia supernovae. We demonstrate how our model can identify previously unobserved kinds of transients and produce a sample that is 90% pure. The ParSNIP model can also estimate distances to Type Ia supernovae in the PS1 data set with an rms of 0.150 ± 0.007 mag compared to 0.155 ± 0.008 mag for the SALT2 model on the same sample. We discuss how our model could be used to produce distance estimates for supernova cosmology without the need for explicit classification.