What can Gaussian processes really tell us about supernova light curves? Consequences for Type II(b) morphologies and genealogies

Abstract

ABSTRACT Machine learning has become widely used in astronomy. Gaussian process (GP) regression in particular has been employed a number of times to fit or resample supernova (SN) light curves, however by their nature typical GP models are not suited to fit SN photometric data and they will be prone to overfitting. Recently GP resampling was used in the context of studying the morphologies of Type II and IIb SNe and they were found to be clearly distinct with respect to four parameters: the rise time (trise), the magnitude difference between 40 and 30 d post-explosion (Δm40–30), the earliest maximum (post-peak) of the first derivative (dm1), and minimum of the second derivative (dm2). Here we take a close look at GP regression and its limitations in the context of SN light curves in general, and we also discuss the uncertainties on these specific parameters, finding that dm1 and dm2 cannot give reliable astrophysical information. We do reproduce the clustering in trise–Δm40–30 space, although it is not as clear cut as previously presented. The best strategy to accurately populate the trise–Δm40–30 space will be to use an expanded sample of high-quality light curves [such as those in the Asteroid Terrestrial-impact Last Alert System (ATLAS) transient survey] and analytical fitting methods. Finally, using the bpass fiducial models, we predict that future photometric studies will reveal clear clustering of the Type IIb and II light curve morphologies with a distinct continuum of transitional events.